New insights for precision treatment of glioblastoma from analysis of single-cell lncRNA expression

Abstract

Introduction

Glioblastoma (GBM) is a complex disease with high intratumoral heterogeneity, understanding the molecular characteristics of intratumoral heterogeneity accurately is the basis for precision treatment. Although the existing typing strategy based on tumor molecular characteristics has a positive effect, there is still room for improvement, which is mainly because the traditional typing is completed based on the sequencing data of tissue samples, that is, the obtained data are the average level of patient tumor tissues, masking the intratumoral heterogeneity of a single patient and cannot reflect the real level of patient tumor cells. At present, cancer molecular typing is mostly performed based on transcriptome (RNA-seq) without considering lncRNA molecules that are also tissue-specific and developmental stage-specific. Therefore, in this study, we used lncRNAs as typing markers and combined single-cell expression profiles to retype glioblastoma, providing new ideas for GBM molecular typing, and further analyzed the shortcomings of traditional therapies at the singlecell level based on typing results and proposed new precise therapeutic insights.

Methods

We downloaded GBM single-cell sequencing data from GSE84465 and performed a series of preprocessing. The intratumoral heterogeneity of patients at the single-cell level was revealed using t-SNE, and the room for improvement of the existing traditional histotyping method was revealed using heat map and density curve. Subsequently, to validate the feasibility of lncRNA typing, we compared the similarities and differences of expression patterns between lncRNAs and mRNAs in GBM cells. Then, we used the R package “Seurat” to perform unsupervised clustering of GBM cells for re-typing and performed a detailed analysis of the biological characteristics of each subtype, including differentially expressed lncRNAs and marker lncRNAs. For validation, we performed survival analysis on GBM tissue data from the TCGA database to reveal the impact of different subtypes on patient survival prognosis. Eventually, based on the results, we screened the therapeutic drugs of each subtype by targeting the downstream regulatory genes of lncRNAs and proposed a new precision therapeutic strategy.

Results

GBM has significant intratumoral heterogeneity at the single-cell level, with more than one traditional subtype highly expressed in each patient, which reflects the shortcomings of traditional histotyping. LncRNAs and mRNAs have similar expression patterns in GBM cells, and the expression coefficient of variation of lncRNAs is higher than that of mRNAs, meaning that lncRNAs will better reflect the intratumoral heterogeneity. GBM was reclassified into four subtypes by unsupervised clustering, with different subtypes having different biological characteristics. Survival analysis showed that patients with different subtype compositions had different prognostic outcomes, so different subtypes had different effects on patient prognosis. Based on this, 47 drugs were screened for treatment. There are both shared and unique drugs between different subtypes. A new precision treatment strategy was proposed: for patients with different subtypes, in addition to the combination of drugs targeting single subtype, drugs targeting multiple subtypes can also be selected.

Conclusion

Intratumoral heterogeneity may lead to poor prognosis or recurrence after treatment, and more precise typing of GBM can be performed based on single-cell lncRNA expression profiles. The biological characteristics possessed by different subtypes will have different effects on patients, such as survival time. For different subtypes, there are both drugs targeting single subtype and drugs targeting multiple subtypes, and we prefer drugs targeting multiple subtypes because this strategy can maximize medication efficiency and reduce the types of medication to reduce risks and side effects.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00432-021-03584-9.

Introduction

Glioblastoma (GBM) is the most common form of malignant brain cancer occurring in the adult population (Otani et al. 2017). Due to its complex heterogeneity, there are large differences in prognostic of glioblastoma patients (Ohgaki and Kleihues (2005)). At present, the commonly recognized molecular typing of GBM in clinical was first proposed by Roel G.W. Verhaak et al. in 2009 (Verhaak et al. 2010). They classified GBM into four subtypes: proneural, neural, classical, and mesenchymal based on genomic and transcriptomic profile of tumor tissues, and revealed that different subtypes responded differently to invasive treatment by clinical survival analysis, with the classical subtype having the best therapeutic effect and the proneural subtype having no therapeutic effect. This highlights the positive role of molecular typing of tumors in guiding clinical precision treatment. However, traditional bulk sequencing methods based on multicellular levels obtain the average expression of multiple cells, which loses heterogeneity information between individual cells (Hwang et al. 2018) and may lose valuable information to guide clinical treatment. Since glioblastoma is a heterogeneous tumor, there is still room for improvement in molecular typing based on tissue sequencing data.

Single-cell sequencing, which uses next-generation DNA sequencing (NGS) technology to analyze DNA sequences of single cells, can obtain the differences of the express status of single cells under specific microenvironment to facilitate the study of their functions (Hedlund and Deng 2018). The advent and the level that tends to mature of single-cell sequencing technology make it possible to apply it in the field of cancer molecular typing.

In addition to this, current molecular typing methods for cancers are usually based on coding gene expression profiles instead of lncRNA which also plays an important role during cell development and proliferation. LncRNA is space-specific and time-specific during cell growth and development, its misexpression is likely to lead to the initiation, growth, and metastasis of cancer cells (Yang et al. 2014), therefore, it also has the potential to be used as a molecular typing marker.

Based on the above, our study precisely analyzed the intratumoral heterogeneity of GBM at the single-cell level that has a higher resolution to reveal the reasons for the significant differences in the treatment of different patients by traditional therapies, retyped GBM using single-cell lncRNA expression profiles, and validated the validity of the new typing using data from TCGA. Finally, we screened the corresponding drugs based on the characteristics of different subtypes and proposed new precise treatment insights: 1. Monotherapy may not target all cancer cells but some specific subtypes, so residual cell subtypes put cancer at risk of recurrence. 2. The emergence of subtypes expressing drug resistance-related markers will make the drug-resistance characteristics appear at the cell population level first and then may produce drug resistance in the late stage of treatment. 3. Medication should be customized for patients that expressed different subtypes and combined medication strategies should be reasonably selected for patients expressing drug resistance-related subtypes to enhance therapeutic sensitivity.

Methods

Data processing

The single-cell expression profiling data used in this study were derived from GSE84465 and contained four glioblastoma patients with a mean age of 54 years. A total of 3589 cells were obtained from tumor tissues, including 1091 cancer cells, and the other cells were non-cancer cells, such as immune cells. Detailed information on patients is shown in Table 1. Cell-type signatures have been given by Spyros Darmanis et al. (Darmanis et al. 2017), so we directly divided all cells into cancer and non-cancer cells according to them to construct expression profiles, respectively. RNAs expressed by less than 1% of the cells were defined as low-expressed RNAs and eliminated. The expression profiles composed of remained RNA were transformed using log2 (X + 1) to normalize. All the data preprocessing process is completed in R.

Table 1.

Sample information

Patient ID	Age	Stage	Cancer cells	Total cells
BT_S1	55	GBM (IV)	268	489
BT_S2	54	GBM (IV)	531	1169
BT_S4	60	GBM (IV)	163	1542
BT_S6	48	GBM (IV)	129	389

Intratumoral heterogeneity analysis

To better reveal the intratumoral heterogeneity of GBM at the single-cell level, we performed t-SNE dimensionality reduction and visualization of expression of cancer cells from all patients. The t-SNE method was implemented using the R package "Rtsne." The dimensionality reduction results were presented as scatter plots using "plot".

At present, the commonly recognized molecular typing of GBM is proposed by Roel G.W. Verhaak et al., who divided GBM into four subtypes. We obtained the marker genes of four traditional subtypes in the relevant literature (Table 2). The traditional subtype characteristics of every single cell were represented by the mean expression of marker genes of each subtype, which were presented using the heat map and density curve by “pheatmap” and “geom_density”.

Table 2.

Marker genes of each subtype

Proneural	Neural	Classical	Mesenchymal
DLL3 NKX2-2 SOX2 OLIG2 PDGFRA	NEFL GABRA1 SYT1 SLC12A5	FGFR3 PDGFA EGFR AKT2 NES	ILR4 TRADD RELB CHI3L1(YKL40) MET CD44 MERTK

Analysis of expression characteristics of lncRNAs in GBM cells

We downloaded a list of currently known human lncRNAs from the Ensembl database (Cunningham et al. 2018) and then intersected with our profile to construct lncRNA expression profile, which totaled 259 lncRNAs.

To show that lncRNAs can be used as the typing marker and have expression heterogeneity, which is the same as traditional marker genes, we first compared the expression patterns from two perspectives: 1. The expression ratio of each lncRNA and gene in cancer cells. 2. The expression proportions of lncRNAs and genes in each cell. Afterwards, to further illustrate that lncRNAs are more suitable as typing markers than genes, the coefficients of variation of the two were compared and plotted in box plots using “geom_boxplot”. The coefficient of variation can reflect the degree of dispersion between data, the higher the coefficient of variation, the higher the cell-to-cell expression heterogeneity.

Re-typing based on single-cell lncRNA expression profiles

Seurat is a method specifically designed and developed for single-cell data mining (Stuart et al. 2019), which encompasses a range of methods from quality control to visual analysis. We performed a detailed analysis of GBM single-cell lncRNA expression profiles based on the Seurat method.

Counts normalization by “method = "LogNormalize" and cell quality control of lncRNA expression profiles was performed first. We removed cells that do not express lncRNAs and performed feature selection using “FindVariableFeatures” on all lncRNAs to screen out lncRNAs with a high variation for the next principal component analysis (PCA). Subsequently, we combined two methods, "JackStraw" and "Elbow", to determine the optimal number of principal components: JackStraw method uses random sampling to select the principal components with a significant correlation. For large data sets, running this method will be relatively slow, and it may not find the appropriate critical point. Therefore, we combine the Elbow method to increase the accuracy of principal component selection. Elbow method ranks the variance explained by each principal component based on the proportion, and it is reasonable to find the inflection point or make the selected principal component include a sufficiently large variance ratio. The use of principal component for downstream unsupervised cluster analysis and visualization can not only greatly improve the operation speed of the program, but also the accuracy. To make the selection of principal components more objective, we selected the first 3, 4, and 5 principal components to determine the optimal number of principal components through downstream clustering and differential expression results, and finally selected the first three principal components.

Unsupervised clustering of all cancer cells was performed using “FindNeighbors” and "FindClusters", and then two-dimensionality reduction visualization methods, UMap (Ultsch and Mörchen 2006) and t-SNE, were used to display the clustering results, respectively, by function “RunUMAP” and “RunTSNE”.

Identification of newly typed differentially expressed and marker lncRNAs

Differentially expressed lncRNAs of each subtype were screened using "FindAllMarkers" and setting default parameters, which looks for differentially expressed lncRNAs by calculating the statistical significance(p < 0.05) between the expression of a subtype and the means of other subtypes (Supplementary Table 5-1). This may lead to the same differentially expressed lncRNAs appears in different subtypes, but it can still reflect the high-expression characteristics of lncRNAs in a certain subtype. The expression of differentially expressed lncRNAs with the min P value and avg_logFC > 0 for each subtype was visualized by three methods given by Seurat: violin plot(function "VlnPlot"), UMap-based expression region display(function "FeaturePlot"), and heatmap(function "DoHeatmap").

To further determine the subtype-specific marker lncRNAs, "get_marker_genes" in the "SC3" package (Kiselev et al. 2017) was used. This method is able to find marker lncRNAs unique to each subtype. P < 0.05 was used as a threshold to screen marker lncRNAs (Supplementary Table 5-2).

Survival Analysis Validation

According to the lncRNA expression profiles of each subtype, the average expression matrices were generated. To validate subtyping results, we downloaded the tissue expression profiles of 152 glioblastoma patient samples with clinical information from the TCGA-GBM project (Tomczak et al. 2015), then the average expression matrices and tissue expression profiles were normalized and centralized, respectively, to eliminate the differences in the expression levels between patients.

Pearson coefficients between each subtype and each patient from TCGA based on the expression of lncRNAs were calculated to construct a correlation coefficient matrix, followed by hierarchical clustering using Ward's minimum variance with Euclidean distance as a metric. Then, survival analysis was performed for groups of patients after clustering.

To investigate the impact of subtype composition on patient prognosis, the Pearson coefficient distribution of each subtype in every patient population was analyzed and was visualized by heatmap.

Identification and screening of candidate drugs

Due to the wide variety of lncRNAs and the current research on lncRNAs is still shallow compared to coding genes, there are few drugs directly targeting lncRNAs. So four databases, Ensembl, ENCORI (Li et al. 2014), LncRNA2Target V2.0 (Cheng et al. 2019), and LncRNADisease v2.0 (Bao et al. 2019) were first used to search the downstream regulatory genes of all marker lncRNAs and then use them as targets to search the drug bank (Wishart et al. 2018) and TDD (Wang et al. 2020) databases for related drugs. Drugs labeled as “approved” in the drug bank database and drugs that corresponding targets labeled as “successful targets” in the TDD database were screened.

The subtype-target-drug network was drawn by cytoscape (Shannon et al. 2003) to visualize the targets of each subtype and the drug sharing relationship, and the Venn diagram was used to further analyze the drug sharing between each subtype.

Results

Intratumoral heterogeneity of GBM

Our study mainly analyzes the intratumoral heterogeneity of GBM, which is generally accepted that has a major impact on the treatment and prognosis of GBM. Cancer cells were isolated according to the cell type label given by Spyros Darmanis et al., which are used to construct cancer cell expression profiles, followed by a series of quality controls and non-linear dimensionality reduction visualization using the t-SNE method (Fig. 1). We mapped multidimensional data into a two-dimensional plane to enable more intuitive observation of intratumoral heterogeneity of GBM on a two-dimensional level. It can be observed in the t-SNE dimensionality reduction scatter diagram that the cancer cells from each patient have different degrees and sizes of outlying cell populations, which show that each patient has significant intratumoral heterogeneity.

Fig. 1.

Scatterplot of dimensionality reduction of t-SNE in cancer cells from each patient. a Patient BT_S1. b Patient BT_S2. c Patient BT_S4. d Patient BT_S6

The above results show the great advantage of revealing the intratumoral heterogeneity of GBM at the single-cell level over traditional tissue analysis, but also illustrate the necessity and feasibility of more precise molecular typing of complex cancer cells at the single-cell resolution.

Shortcomings of traditional molecular typing method

Traditional typing methods analyze the transcriptome expression of cancer tissues, but GBM has significant intratumoral heterogeneity, so there may be potential shortcomings in investigating its transcriptional expression characteristics from the tissue level for cancer typing.

Roel G.W. Verhaak et al. determined marker genes for the current GBM typing and verified its effectiveness. The mean expression of marker genes of each subtype was used as the traditional subtype characteristic corresponding to a certain cell. As shown in Fig. 2, the cancer cells of patient S1 mainly expressed the characteristics of classical and proneural subtypes, S2 mainly expressed the characteristics of classical, mesenchymal, and proneural subtypes, and the expression of S4 and S6 was similar to that of S2, but the degree of expression of each subtype was different between patients. The cancer cells of each patient express at least two traditional subtype characteristics, which shows that from the single-cell level, the cancer tissues of each patient do not completely belong to a single subtype, while the tissue expression profile used for traditional typing is similar to the average level of single cells of cancer tissues, so it is possible to mask the intratumoral heterogeneity existing in individual patients due to the dominant number of cells in a certain subtype, which is detrimental to the treatment and prognosis of patients.

Fig. 2.

Expression of traditional markers in each patient. a Patient BT_S1. b Patient BT_S2. c Patient BT_S4. d Patient BT_S6

Another way, the density curve (Fig. 3), shows the expression of different subtype marker genes in each patient. Since it is shown in the heatmap that all patients express little Neural characteristics, the density curves only compare the other three. The curve showed that multiple high-expression peaks occurred in each patient, patient S1 mainly expressed two subtype characteristics, S2, S4, and S6 expressed three subtype characteristics, which were consistent with the heatmap results.

Fig. 3.

Density curve of cells expressing traditional markers. a Patient BT_S1. b Patient BT_S2. c Patient BT_S4. d Patient BT_S6

Expression characteristics of lncRNAs in GBM cells

Since the tissue and developmental stage-specificity of lncRNAs makes them likely to be used as typing markers, and lncRNAs play an important role in tumor proliferation, cell invasion, migration, apoptosis, and communication, we explored the feasibility of lncRNAs as a new typing basis firstly.

We performed a series of preprocessing of the above lncRNAs, including log normalization and quality control, and finally retained 71 lncRNAs. The expression patterns of these lncRNAs and known marker genes in cancer cells were first compared and depicted by density curves (Fig. 4). We mainly compared the expression patterns of the two from two perspectives: 1. The proportion of cells expressing each lncRNA and marker gene, the vast majority of lncRNAs are zero expressed in most cancer cells, and a few lncRNAs are expressed in some cancer cells, which is approximately the same as the expression pattern of marker genes. 2. The expression of lncRNAs and marker genes in each cancer cell, only a small amount of specific lncRNAs and marker genes were expressed in cancer cells, and the remaining lncRNAs and marker genes were in a low or no expression state. In summary, the expression patterns of lncRNAs and marker genes in GBM cells are almost the same, so lncRNAs also have expression heterogeneity in cancer cells and can be used as a basis to assess the overall expression of cancer cells, which provides a prerequisite for our subsequent retyping of GBM.

Fig. 4.

Comparison of expression patterns of traditional marker and lncRNA. a Cell expression proportion of each lncRNA. b Cell expression proportion of each marker gene. c Expression density of lncRNAs in each cell. d Expression density of marker genes in each cell

Subsequently, to further illustrate that lncRNAs are more suitable as a basis for molecular typing than marker genes, we compared the expression heterogeneity of lncRNAs and marker genes in GBM cancer cells by the variation coefficients of expression (Fig. 5). The coefficient of variation can reflect the degree of dispersion between data, so the higher the coefficient of variation, the higher the cell-to-cell expression heterogeneity. Box plots show that the coefficient of variation of lncRNAs is significantly higher than that of marker genes, which shows the ability of lncRNAs to representing intercellular dispersion is better.

Fig. 5.

Comparison of the coefficient of variation of lncRNAs and markers

Through the comparison of the above two aspects, we found that lncRNAs are not only able to characterize intratumoral heterogeneity, but also would have more representative than marker genes, which is more suitable for use as GBM molecular typing markers.

Reclassification of GBM based on single-cell lncRNA expression profiling

Seurat object was created using raw expression profiles of 259 lncRNAs from 1091 cancer cells. LncRNAs in this expression profile were first analyzed from two perspectives to perform quality control (Fig. 6): 1. The total expression value of lncRNAs in each cell 2. The number of lncRNAs expressed in each cell. The total expression value of lncRNAs in most of the cells was located around 5000 counts, and at the same time, the number of lncRNAs expressed by the cells was concentrated around 10. In addition to this, there are a few cells that do not express any lncRNAs, and we consider that these cells were not available in this study, so the cells that do not express any lncRNAs were deleted. Finally, 1083 screened cells were used for the next analysis.

Fig. 6.

Expression state of lncRNAs in cancer cells. a Total expression counts of lncRNAs per cell. b The number of lncRNAs expressed per cell

Subsequently, all lncRNAs were subjected to feature selection to find out lncRNAs with high variation properties. High variation lncRNAs were extracted from a total of 259 lncRNAs that will play a crucial role in PCA analysis. The data were scaled using "ScaleData". Data scaling enables the expression value of each gene to be changed so that the average expression between cells is 0 and the variance is 1, then lncRNAs can be given equal weight at downstream analysis so that highly expressed lncRNAs will not be absolutely dominant.

Principal component analysis (PCA) based on high variant lncRNA expression profiles was performed to compress the data, speed up downstream analysis, and improve accuracy. The most critical step of PCA is to determine the optimal number of principal components, that is, the dimensionality of the data after dimensionality reduction. Seurat gives two auxiliary methods for selecting the best dimension number: “JackStraw” (represented as J) and “Elbow” (represented as E). We used the above two methods to analyze after PCA (Fig. 7), J method showed that from principal component 2 to principal component 3, the score increased significantly, indicating that the importance decreased rapidly, E method showed that there was a significant inflection point from principal component 3 to principal component 4, so the first three principal components were selected for downstream analysis in combination with the two methods.

Fig. 7.

Auxiliary methods for principal component selection. a Method JackStraw. b Method Elbow

Unsupervised clustering based on the first three principal components was performed using "FindClusters". The total number of cell clusters is 4, and the clustering result was visualized by two non-linear dimensionality reduction visualization methods, UMap and t-SNE (Fig. 8). Cancer cells after unsupervised clustering based on lncRNA expression profiles were able to be clearly displayed as four cell clusters from both UMap and t-SNE dimensionality reduction scatter plots.

Fig. 8.

Cluster results. a Method UMAP. b Method t-SNE

Differentially expressed lncRNAs of each subtype

Seurat can also be used to screening differentially expressed lncRNAs in each subtype, and the expression state of the most significant lncRNAs was displayed by the violin diagram (Fig. 9), Umap dimensionality reduction-based expression region (Fig. 10), and heat-diagram (Fig. 11).

Fig. 9.

Violin plots of differentially expressed lncRNAs

Fig. 10.

Visualizing the spatial distribution of differentially expressed lncRNAs based on UMAP

Fig. 11.

Heatmap of differentially expressed lncRNAs

Through the visualization of the three methods, it can be observed that differentially expressed lncRNAs have high-expression characteristics in their corresponding subtypes. Highly expressed regions of differentially expressed lncRNAs of each subtype have a dislocation distribution in UMAP and almost along diagonal in the heatmap. Therefore, they can effectively reflect the biological characteristics of each subtype.

Survival differences among subtypes

To validate lncRNA-based typing and further analyze the effect of different subtypes on patient prognosis, the expression profile of 152 patients with survival information was downloaded from TCGA, and the patients were grouped by hierarchical clustering based on the Pearson coefficient matrix between patients and each subtype. Survival analysis was performed on the three patient populations to compare the survival status differences between them (Fig. 12).

Fig. 12.

Survival curve of three patient populations. The red curve represents patient population 1, the green represents patient population 2, and the blue represents patient population 3

The prognosis of patient population 2 was significantly worse than that of population 1 and population 3. The survival time of population 1 was better than that of population 2 on the whole. We found by calculation that the mean age of patient population 3 with the best prognosis was 57.5 years, followed by 60 years in population 1, and 65.2 years in population 2. Generally, the patient's age increased, the treatment difficulty increased, so the survival status was poor.

To further analyze the effect of each subtype on the survival time of patients, we plotted the heatmap of the Pearson correlation coefficient between patients and each subtype (Fig. 13). In the correlation heatmap, the redder the color, the stronger the positive correlation, indicating that the patient's tissue expression is closer to this subtype. The tissue expression level is an approximate average expression of the cells in this tissue and is dominated by the main cell subtypes, so the correlation can also reflect the proportion of cell subtypes in the patient tissue. It is inferred that the main subtype composition of patient population 3 with the best prognosis is 2, 3, and 4, and that of population 1 is 2 and 3. Subtype 3 expression is the most significant in population 2. Besides, subtype 1 is expressed in patient population 2 compared with the other two populations.

Fig. 13.

Heat map of Pearson correlation matrix. There are four molecular types on the horizontal axis and individual patients on the vertical axis, and the color of each cell is determined by the Pearson correlation coefficient between this patient and the corresponding subtype

Patient population 3 was distinguished from the other two populations by the expression of isoform 4, whose marker lncRNA was MEG3 (p = 1.29E-114). It has been supported in the literature that high expression of MEG3 can significantly inhibit the proliferation of glioma cells and can promote both their apoptotic and autophagic processes. The expression level of MEG3 showed a negative relationship with the WHO grade of the patients and a positive correlation with the Karnofsky score (Zhao et al. 2018). In clinical treatment, increasing the expression level of MEG3 is also able to enhance the chemosensitivity of the U87 cell line to cisplatin so that enhancing the therapeutic efficacy (Ma et al. 2017). Among the three patient populations, only population three characterized cell subtype four expressing MEG3, and the prognosis was also relatively best.

Patient population one predominantly expressed cell subtypes 2, 3 with a marker of LINC00461 (p = 8.74E-15) and LINC00339 (p = 2.06E-25), respectively. LINC00461 can affect cell proliferation, migration, and invasion through MAPK/ERK, PI3K/AKT, and possibly other signaling pathways, promoting the development of glioma (Yang et al. 2017). The effect of LINC00339 in glioma is similar to LINC00461, which also promotes the development of glioma. It has been found that the expression level of LINC00339 is closely related to the anti-tumor effect of the drug avasimibe. Experiments have shown that avasimibe inhibits the proliferation, migration, and invasion of glioma cell lines by inhibiting LINC00339 in vitro and in vivo, which provides a new way for glioma therapy (Luo et al. 2020).

Patient population two predominantly expressed cell subtype 3, but the expression level of cell subtype 1 was improved compared with the other two patient populations. The marker lncRNA of subtype 1 is MALAT1 (p = 5.77E-25), which is closely related to the resistance of chemotherapeutic drugs by a variety of studies, and MALAT1 can increase drug resistance of cancer by inhibiting miR-203 and promoting the expression of thymidylate synthase (TS) (Chen et al. 2017). Silencing of MALAT1 can significantly improve the sensitivity of glioma cells to chemotherapeutic drugs, including temozolomide, the current first-line chemotherapeutic drug for the treatment of GBM (Kim et al. 2018). Therefore, the reason for the worst prognosis in this patient population may be the emergence of cell subtype 1 accompanied by the expression of its marker lncRNA MALAT1, which improves the drug resistance of patients and increases the difficulty of treatment.

The analysis showed that with the development of glioblastoma, the cellular subtypes composed of patient tissues were intensively expressed from 2, 3, 4 to 2, 3, then to 3, and the carcinogenicity gradually increased. But the heterogeneity did not disappear, and each patient population still expressed two or more subtype characteristics. The disappearance of subtype 4 and the emergence of subtype 1 accompanied by decreased expression of MEG3 and increased expression of MALAT1, reduce the suppression of cancer and increase the drug resistance, which leads to a gradual deterioration of survival prognosis.

Candidate drugs based on the new classification

Cancer typing at single-cell resolution can identify the intratumoral heterogeneity that cannot be presented by traditional tissue molecular typing. A patient's tumor tissue has potential multiple subtypes and the proportion of each subtype is different. The histotype method reflects the average level of all cells, thus resulting in the subtype that has a huge-proportion dominants, the typing characteristics of tissue cells. Clinical medication often ignores the secondary subtype and is only for the main, which may lead to poor therapeutic effect, gradually increasing drug resistance of tissue during the screening of subtype by the drug, and ultimately presenting consequences of poor prognosis.

To further provide a reference for clinical medication, we identified and screened a total of 47 drugs based on the marker lncRNAs of each subtype, and the subtype-target-drug relationship table (Supplementary Table 7-1) showed that there were different degrees of target sharing or drug sharing among each subtype.

The Venn diagram (Fig. 14) provides a more intuitive perspective of the number of drugs shared between subtypes:

Fig. 14.

The number of drugs shared among subtypes

Subtypes 1 and 3 shared the largest number of drugs, 8, which were: Arsenic trioxide, Colchicine, Podofilox, Vinblastine, Vinorelbine, Vincristine, Vinflunine, and Vindesine. The family of Vinblastine accounted for the majority of these drugs, and Vinblastine can significantly reduce the resistance of glioblastoma to temozolomide (Kipper et al. 2018). Besides, Arsenic trioxide has also been reported to be able to be used in combination with temozolomide to play a better therapeutic effect in glioblastoma (Bureta et al. 2019). Previous analyses have shown that subtype 1 mainly expresses marker lncRNAs associated with temozolomide resistance which leads to poor response to traditional temozolomide monotherapy, and the patient population expressing this subtype has the worst prognosis, so the use of Vinblastine and Arsenic trioxide is essential for the treatment of such patients. Colchicine is an anti-microtubule and anti-mitotic drug, and its derivative AD1 has the potential to be used as a chemotherapeutic drug to inhibit glioma development (Fang et al. 2015). Podofilox is currently mainly used to treat condyloma, and its derivatives have anti-cancer properties, too (Zhang et al. 2018).

The subtypes 1 and 2 shared drug is Acyclovir, which is a nucleoside analogue mainly used to treat viral infections and may be a potential therapeutic drug for glioblastoma (Ozdemir and Gokturk 2019). The shared drugs of subtypes 2 and 3 are Fludeoxyglucose (18F) and Moxetumomab Pasudotox, and subtypes 2 and 4 are Omacetaxine mepesuccinate. The 3 and 4 subtype-shared drugs were Ivosidenib and Indocyanine green acid form, i.e. among them, Ivosidenib is an IDH1 inhibitor, while IDH1 mutation is one of the typical features of glioblastoma, so it has an important therapeutic effect during the formation and development of glioblastoma (Dhillon 2018).

There are both exclusive drugs and shared drugs between different subtypes. In clinical settings, drugs for multiple subtypes should be selected for treatment as far as possible. Compared with the traditional combination strategy of different drugs for different subtypes, the selection of drugs targeting multiple subtypes for treatment can significantly reduce the number of drugs for use, maximize the efficiency of medication while also minimize the toxicity and risk caused by drugs. For example, in three patient populations from TCGA, population 1 expresses both subtypes 2 and 3, but traditional typing methods are limited to tissue expression levels, so it is possible to use the drug only for the main subtypes expressed in this population. Single-cell typing methods can more effectively reflect intratumoral heterogeneity. Drug Moxetumomab Pasudotox that targets EEF2 (Leprivier et al. 2013) and is included in the candidate drug list of both subtype 2 and 3 could be selected for the treatment in patient population 1.

In addition, subtype 1 expresses marker lncRNAs related to drug resistance. Subtypes 1 and 3 are expressed in older patients and have a poor prognosis. Therefore, in the treatment for those patients, not only the therapeutic effect of the drug itself on the tumor should be considered, but also the drug sensitivity should be improved. For example, in the treatment of patient population two expressing subtype 1, the combination of vinblastine and temozolomide can significantly improve the effect of tumor resistance, thereby enhancing the killing effect on tumors.

Discussion

The accurate classification of glioblastoma has extremely important reference value for its diagnosis, treatment, and prognosis. Traditional typing methods are based on tissue transcriptome expression profiling and use coding genes as markers for molecular typing. In our study, we explored a new method for molecular typing of glioblastoma, that is, typing based on lncRNA expression profiles and combining with currently rapidly developing single-cell sequencing technologies.

Significant heterogeneity in glioblastoma was found by expression analysis of single cancer cells in each patient. We found that more than one traditional subtype was expressed in all patients at the single-cell level. This intratumoral heterogeneity is likely to lead to poor prognosis as well as cancer recurrence after treatment. These can also reflect the significant superiority of single-cell technology over traditional tissue analysis. In GBM cells, the expression patterns of lncRNAs and genes have both similarities and differences. They are all characterized by a little expression, but our experiments demonstrate that lncRNAs can better reflect intercellular heterogeneity than genes. Four subtypes were revealed by retyping GBM based on single-cell lncRNA expression profiles. These four subtypes have different biological properties and are characterized by their respective differentially expressed lncRNAs. Validation by TCGA glioblastoma data revealed that there were significant differences in the survival status of patient populations with different subtype composition. The weakening of cancer suppression, enhanced carcinogenesis, and the emergence of drug-resistance characterization are all the reasons for the gradual deterioration of patient prognosis. Finally, based on our results, we screened marketed drugs for different subtypes or multiple subtypes and proposed novel medication strategies.

Precision treatment of tumors is currently one of the central problems in the field of precision medicine. Our study provides a molecular retyping of glioblastoma based on the analysis of single-cell lncRNA expression profiles in glioblastoma and presents new precise therapeutic insights into glioblastoma based on this. Through intratumoral heterogeneity analysis, we can see the great advantages of single-cell sequencing technology over traditional tissue sequencing, and we believe that with the gradual maturation and development of single-cell technology, there will be more and more room for its application in the field of cancer research. On the other hand, lncRNAs, as molecules that play an important role in the entire gene expression process, will be studied more and more thoroughly about their types and mechanisms. The breakthrough in lncRNA research will also promote the development of cancer research.

Our study is an attempt and a small step to combine single-cell technology and lncRNAs for cancer typing at one time. The disadvantage of this study is that the amount of data we use is limited, but it still provides a feasible new idea for researchers.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

This work was supported by the National Natural Science Foundation of China [Grant No. 61671191].

Declarations

Conflict of interest

The authors declare no competing financial interest.