Integration of RNA-Seq and proteomics data identifies glioblastoma multiforme surfaceome signature

The analysis combined the TCGA-GBM and GTEx normal brain RNA-Seq read count data. The GBM RNA-Seq gene raw read counts from TCGA were downloaded from Genomics Data Commons Data Portal (https://​portal.​gdc.​cancer.​gov). GTEx data were used for the normal brain tissues. The GTEx data used for the analyses described in this manuscript were obtained from the GTEx Portal on 29/03/19. We downloaded RNA-Seq gene raw read counts (from the cortex, frontal cortex, anterior cingulate cortex) from the GTEx portal (https://​gtexportal.​org/​home/​datasets). This allows us to perform the analysis of the differentially expressed gene on the 166 samples of GBM tumour from TCGA and 408 samples of normal brain tissues data from GTEx. The RNA-Seq raw read counts pre-processing steps involve are data filtering and data normalization. The normalization process of both data set was then performed by using mean as gene-level normalization using log₂-counts per million where raw data are adjusted to account for factors that will prevent direct comparison of expression measures and to safeguard the expression distributions are similar for each sample across the whole experiment. Data that unlikely to be informative or simply erroneous data will be removed by using variance filter (less than 15) and low abundance (less than 4).

The identified differentially expressed genes (DEGs) of glioblastoma were classified into cell-surface genes set as discussed in the main text (See Results 2.4). The classification of the gene sets was performed based on the mapping set of DEGs with this resource. Other genes, which did not map to this resource were removed from the final dataset.

DEGs analysis was performed using NetworkAnalyst [14], a web-based application tool for visualizing molecular and entity interactions. This platform utilizes the statistical method on data comparison from the R package, limma to identify genes whose expression is different. Genes that have adjusted p-value < 0.05 and log2 fold change |2| were considered as statistically significant DEGs.

The enrichment analysis of the identified glioblastoma associated genes was performed using DAVID (https://​david.​ncifcrf.​gov/​), a web-based tool for analyzing functional gene analysis. The tool comprises databases from various public resources for biological analysis. The enrichment analysis such as GO and KEGG pathways were performed with top results as per gene counts.

A biological database for known and predicted protein-protein interactions called IMEx interactome database (https://​www.​imexconsortium.​org) was used to construct the protein-protein interaction (PPI) of the DEGs. The network of interacting proteins was extracted and visualized using NetworkAnalyst. The top 87-gene modules of highly interacting gene clusters among the DEG were found with default parameters. For the classified gene sets, the PPI network was constructed and the network topological parameters i.e. degree and betweenness centrality were calculated.

Co-expression analysis was performed using Graphia Professional (https://​kajeka.​com/​graphia-professional/​), previously known as BioLayout Express^3D [15] using raw read counts and then saved as an “.expression” file. This contains a unique identifier for each row of data. Following import into Graphia Professional, a pairwise Pearson correlation matrix was calculated thereby performing a gene vs. gene comparison of the expression profile of each gene. All Pearson correlations where r > 0.7 were saved to a “.pearson” file. Based on a user-defined threshold of r > 0.75, an undirected network graph of the data was generated. In this context, nodes represent individual genes and the edges between them represent Pearson correlation coefficients above the selected threshold (r > 0.75). CD44 was selected along with its neighbour in the network, representing CD44 co-expression partners. The class set of CD44 co-expressed genes were visualized to compare the expression values in this class set with genes in normal samples.

We utilized the publicly available TCGA and GTEx RNA-Seq database as our primary sources of GBM tumour and normal brain tissue transcriptomic data, respectively. We downloaded the datasets containing RNA-Seq gene expression profiles and clinical information of 166 patients from TCGA-GBM and 408 normal brain tissues from the GTEx database. The combined data were stratified based on sex, age and treatment as shown in Table 1. Out of a total of 166 GBM cases, 104 cases (62.7%) were male, 56 cases (33.7%) were female and 6 cases did not have sex specification. GBM is more prevalent in patients aged ≥60 years old which accounts for 42.8% of total cases in the TCGA GBM cohort. Fifty-two patients (31.3%) have undergone treatments whereas 62.1% of cases did not have any treatment data. Unfortunately, the clinical data for the GTEx normal brain samples are not publicly available.

The analysis pipeline employed in this study is depicted in Fig. 1. Briefly, the RNA-Seq raw read counts from the two large compendiums, TCGA and GTEx were utilized to identify the differentially expressed genes between GBM and normal brain tissues. Since most GBM cases are generally found in the supratentorial region of the brain such as the cerebral hemisphere [16], we only extracted the RNA-Seq profiles of this region namely the cortex, frontal cortex, anterior cingulate cortex as per GTEx description. We performed t-distributed stochastic neighbour embedding (t-SNE) analysis to reflect the directionality of transcripts expression among GBM tumour and normal brain tissues read count values. The t-SNE plot showed that all RNA-Seq profiles of all GTEx cortex regions clustered together while the GBM RNA-Seq profiles form a separate cluster, thus confirming distinct expression patterns between these groups (Fig. 2A). In total, RNA expression data from 18,021 genes were obtained from these combined TCGA and GTEx datasets but only 13,548 genes passed the quality control check. By applying the cut-off criteria log₂ fold change |2| and adjusted p-value < 0.05, we identified 2381 genes as significantly differentially expressed genes (DEGs) in GBM, of which 648 genes were upregulated and 1733 genes were downregulated (Fig. 2B). The detailed information of the differential gene expression analysis is listed in Supplementary Table S1.

The significant DEGs were then subjected to functional enrichment analysis using Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) tools to define their properties and putative biological relevance in GBM. Interestingly, the GO cellular component analysis of both upregulated and downregulated DEGs showed enrichment of cell surface and membrane-associated proteins (Supplementary Fig. S1A and B). The KEGG pathway enrichment indicated that the upregulated DEGs are involved in pathways related to infectious diseases, pathways in cancer and cell adhesion (Supplementary Fig. S1C). Downregulated genes mainly involve in neuroactive ligand-receptor interaction and major cellular signalling pathways (Supplementary Fig. S1D).

The DEGs were then further filtered and classified into the surfaceome gene set as previously defined by Bausch-Fluck et al. [13], Cunha et al. [17] and Lee et al. [18]. These studies utilized different criteria and stringency in curating the surfaceome gene list. From the overall DEGs in GBM, we identified 395 common cell surface genes within these three surfaceome definitions, including 124 upregulated and 271 downregulated genes (Supplementary Fig. S2A and Supplementary Table S2). We further classified the surfaceome according to their main subclasses, which are receptors, transporters, enzymes, miscellaneous and unclassified, as previously reported by Almén et al. [19]. Among the defined surfaceome subclasses, 42.8% of the significant differentially expressed surfaceome in GBM belong to the receptor subclass (Supplementary Fig. S2B). KEGG analysis of the GBM-enriched cell surface proteins identified pathways related to immune defence and infectious disease pathways while GBM-deficient cell surface genes are enriched in pathways related to neuroactive ligand-receptor interaction and major cellular signalling pathways (Supplementary Fig. S3A and B). These findings are almost similar to the enrichment analysis of overall DEGs in GBM (Supplementary Fig. S1C and D) suggesting that surfaceome has significant roles in dictating GBM cellular activities.

Thus far, we have (i) classified the overall DEGs in GBM using transcriptomics data and (ii) highlighted the differentially expressed cell-surface genes in GBM. Even though this transcriptomics analysis is very informative for biomarker discovery, we aimed to add another layer of analysis to select a more high-confidence cell surface signature for GBM. To attain this, we integrated our transcriptomics analysis data with the publicly available proteomics data. This integration will validate the cell surface genes prediction and eliminate the possible discrepancy between the expression levels of mRNAs and proteins due to post-transcriptional and post-translational modifications. Thus, we gathered the publicly available quantitative mass spectrometry analysis data for both GBM tissues and cell lines. We postulated that GBM tissues and cell lines might have different cell surface repertoires and therefore it is important to stratify between these two sources. Additionally, GBM cell lines cell surface signature, as identified in this present study, could be validated experimentally in future functional studies.

Mass spectrometry analysis of five GBM cell lines revealed the upregulation of EGFR, CD44, PTPRJ, SLC1A5, F2R, and TSPAN6 proteins in these samples [12], whereby the expression level of these proteins was in concordance with our transcriptomics data analysis (Fig. 3). For tissue proteomics, we found several studies that performed comparative GBM vs. normal brain tissues proteome profiling [11, 20‐23]. However, some of these studies either identified only a limited number of proteins or the data are not downloadable. Only one study by Polisetty et al. has identified a large number of proteins in their proteome profiling study that included 1834 high-confidence membrane proteins with more than 2-fold change [11]. We, therefore, used this dataset where we performed an integrative analysis with our analyzed transcriptomics data and identified 10 overlapped genes, MRC2, FCGR3A, HLA-DRA, CD44, CD74, MSR1, CD163, EGFR, ITGB2, PTPRZ1 (Fig. 3). The mRNA expression levels correlated with the protein expression levels except for the PTPRZ1 where the mRNA levels showed upregulation while proteomics data showed downregulation (Supplementary Table S3 and S4). In total, there are 14 genes from the combined tissues and cell lines proteomics that overlapped with our transcriptomics data (Fig. 3). It is important to note that proteins identification in mass spectrometry can be limiting due to protein isolation methods, proteins solubility, and other intrinsic variations that affect the proteins abundance as well as the sensitivity and detection capability of the MS instrumentation [24, 25]. Thus, these limitations may underestimate the results of transcriptomics prediction and proteomics discoveries.

We set out to further analyze the GBM-enriched cell surface markers using protein-protein interaction (PPI) network analysis. This is to better understand the interplay between the cell surface genes within the identified DEGs as well as with other genes. More importantly, this would enable us to further select the genes that are highly interconnected from the integrated proteomics and transcriptomics analysis. Network analysis of the identified differentially expressed cell surface protein genes was performed using NetworkAnalyst [14] to determine the relationship between genes according to the network topological parameters such as degree and betweenness. These parameters reflect the role and property of proteins within the network. The nodes and edges in the PPI network represent the proteins and their interactions, respectively. The GBM-enriched cell surface proteins network contains 1321 nodes and 1767 edges interactions based on a number of validated features including functional experiments, co-expression analysis, text mining, neighbourhood, gene fusion and databases (Fig. 4A). We identified 87-gene modules of clusters and the top cluster genes with more than 30 interactions include VCAM1, EGFR, TGFBR1, CD44, NGFR, ITGB2, DCC, PTPRJ, ANBCA1, HLA-DRA, CCR5 and CSF1R (Fig. 4A and Supplementary Table S5). Vascular Cell Adhesion Molecule 1 (VCAM1) has the highest interacting cluster as it was found to have 426 degrees with a 422,712.18 betweenness score. We subsequently mapped the 14 genes identified from the integrated transcriptomics and proteomics data analysis (Fig. 3) with the top genes that have at least 20 interactions from the PPI network analysis. We found 6 genes that were in common between these two datasets which represent the high-confidence GBM predictive surfaceome markers (Fig. 4B). It is important to highlight that our analysis thus far integrated multi-OMICS data from bulk samples. It is known that GBM suffers from inter and intra-tumoural heterogeneities that contribute to the emergence of several molecular subtypes [26‐28]. Single-cell RNA-sequencing (scRNA-seq) corroborated that the co-existence and interaction between different cells population within the GBM microenvironment drive the GBM cells pro-oncogenic cellular programs. To examine this, we integrated our analysis with the scRNA-seq data containing 24,131 cells from adult and pediatric GBM patients [26]. The study further stratified the cells within the GBM microenvironment into macrophages, oligodendrocytes, T-cells and malignant cells (Supplementary Fig. S4). Of the identified 6 high-confidence cell surface markers, only EGFR was strongly expressed in the malignant GBM cells, while the other genes were strongly expressed in the macrophages (Fig. 4C).

Next, we validated the expression profiles of the identified 6 high-confidence cell surface markers using an independent database, Gene Expression Profiling Interactive Analysis (GEPIA) [29]. GEPIA also combines the TCGA and GTEx gene expression data that were processed from raw reads count and unified using its own pipeline. In line with our findings, the identified GBM cell surface signature genes were confirmed to be significantly upregulated in the GBM GEPIA database (Supplementary Fig. S5A – F). To investigate whether the expression level of these signature genes would modulate/influence GBM patients’ prognosis, we first performed the overall survival analyses on GBM patients who had high or low expression of each of these 6 genes (Supplementary Fig. S6A – F).

However, there were no significant differences in the overall survival between patients who had high or low expression of these 6 individual prioritized genes. Since GBM patients have a low overall survival rate (average < 2 years’ survival post-diagnosis), we postulated that it would be more appropriate to look at the disease-free survival endpoint rather than the overall survival. Moreover, the overall survival endpoint is more suited for a longer follow-up period (typically 5 years) for the data to be meaningful [30]. Hence, we examined the disease-free survival profile of the GBM patients in a similar fashion. We found that high expression of CD44, PTPRJ and HLA-DRA were significantly correlated (p < 0.05) with poor disease-free survival in GBM patients (Supplementary Fig. S7A – S7F).

In addition to performing survival analysis on the individual gene, we also assessed whether combining the level of all 6 GBM signature genes as a group could predict the GBM patients’ overall survival and disease-free survival. We observed that there was no statistically significant difference in the overall survival and disease-free survival between patients who had high expression and low expression of the signature group (Supplementary Fig. S8A – B). Interestingly, by combining only CD44, PTPRJ and HLA-DRA in the gene signature, we found that subjects with high expression of this signature group had significantly poor disease-free survival (p < 0.0084) compared to patients who had low expression of these genes (Supplementary Fig. S9). However, there was still no significant difference in the overall survival between GBM patients in this signature group (Supplementary Fig. S9).

CD44 is a transmembrane receptor and has multifaceted functions in both normal and disease physiology. OMICS studies have identified CD44 to be overexpressed in many types of cancer including glioblastoma [31, 32]. Based on our analysis, CD44 seems particularly important as it can be both identified in transcriptomics and proteomics-based approaches, among the top hub gene and whose high expression correlated with poor disease-free survival. In addition, scRNA-seq identified CD44 was enriched in mesenchymal-like cell state in GBM population, and orthotopic xenografts of these CD44-enriched fractions in immunocompromised mice was able to initiate GBM [26]. We performed a co-expression network analysis to further interrogate its association with other genes using our transcriptomics. The nodes represent in the network analysis represent genes, while the edges represent Pearson correlation above r > 0.75. The neighbouring genes connected to CD44 was extracted and shown in Fig. 5A. There are 27 genes in this complex connected to CD44. Among the highly correlated genes are ELK3, CLIC4, GALNT2, TNC, and VIM. All genes in this CD44 co-expression cluster are highly expressed in GBM compared to normal brain samples (Fig. 5B), further corroborating the biological relevance of CD44 in supporting GBM pathogenesis.

We next determined whether there are any clinically approved drugs targeting the identified high confidence GBM cell surface markers (Supplementary Fig. S4) and components of the constructed CD44 co-expression network (Fig. 5A). To achieve this objective, we utilized the Drugbank database (https://​go.​drugbank.​com/​) and our searches yielded several approved drugs that can be potentially effective or repurposed to target CD44, EGFR, C1R, CALR and TNFSR1A (Table 2). Hyaluronic acid, for example, is a clinically approved ligand for CD44 and this drug has been administered in the clinic to treat diseases such as osteoarthritis [33]. Excessive hyaluronic acid administration has been demonstrated to inhibit tumour growth, possibly by impeding cell-cell interaction [34]. Besides, the use of nanomaterials to enhance the efficiency of hyaluronic acid delivery for cancer therapy is also actively being explored [35, 36]. Thus, the promising features of hyaluronic acid in mediating enhanced drugs or genes delivery to cancer cells via the overexpressed CD44 receptor could potentially be applied and developed for novel GBM therapeutic strategies. In regards to EGFR, several inhibitors and monoclonal antibodies have already been therapeutically approved to target this protein due to its roles as an important driver of tumorigenesis in many cancer types [37].

Moreover, of the 28 components of the CD44 co-expression network (Fig. 5A), only C1R, CALR, and TNFSR1A have drugs that can modulate them (Table 2). For instance, 3 drugs can be used or repurposed to target C1R. The pharmacological activity of Palivizumab to bind the C1R subcomponent is under investigation, whereas the conestat alfa and human C1-esterase inhibitor can directly target the C1R subcomponent and disrupt the complement system activation.