Immune landscape-based machine-learning-assisted subclassification, prognosis, and immunotherapy prediction for glioblastoma

Abstract

Introduction: As a malignant brain tumor, glioblastoma (GBM) is characterized by intratumor heterogeneity, a worse prognosis, and highly invasive, lethal, and refractory natures. Immunotherapy has been becoming a promising strategy to treat diverse cancers. It has been known that there are highly heterogeneous immunosuppressive microenvironments among different GBM molecular subtypes that mainly include classical (CL), mesenchymal (MES), and proneural (PN), respectively. Therefore, an in-depth understanding of immune landscapes among them is essential for identifying novel immune markers of GBM.

Methods and results: In the present study, based on collecting the largest number of 109 immune signatures, we aim to achieve a precise diagnosis, prognosis, and immunotherapy prediction for GBM by performing a comprehensive immunogenomic analysis. Firstly, machine-learning (ML) methods were proposed to evaluate the diagnostic values of these immune signatures, and the optimal classifier was constructed for accurate recognition of three GBM subtypes with robust and promising performance. The prognostic values of these signatures were then confirmed, and a risk score was established to divide all GBM patients into high-, medium-, and low-risk groups with a high predictive accuracy for overall survival (OS). Therefore, complete differential analysis across GBM subtypes was performed in terms of the immune characteristics along with clinicopathological and molecular features, which indicates that MES shows much higher immune heterogeneity compared to CL and PN but has significantly better immunotherapy responses, although MES patients may have an immunosuppressive microenvironment and be more proinflammatory and invasive. Finally, the MES subtype is proved to be more sensitive to 17-AAG, docetaxel, and erlotinib using drug sensitivity analysis and three compounds of AS-703026, PD-0325901, and MEK1-2-inhibitor might be potential therapeutic agents.

Conclusion: Overall, the findings of this research could help enhance our understanding of the tumor immune microenvironment and provide new insights for improving the prognosis and immunotherapy of GBM patients.

Keywords: glioblastoma (GBM); immune landscape; immunotherapy; machine-learning (ML); prognosis; subclassification.

Figure 1

(A) Performance comparisons of different ML models based on 73 immune signatures retained at least 50 times in SVM-RFE 100 times. The ACC and the F1 values of each model are the averages of those for 50 different testing sets generated by randomly dividing each subtype dataset into training and testing sets 50 times. (B) Comparisons of RBF-based SVM models based on five feature subsets containing immune signatures retained at least 50, 70, 80, 90, and 100 times, respectively. The ACC and the F1 values of each model are the averages of those for 50 different testing sets generated by randomly dividing each subtype dataset into training and testing sets 50 times. (C) The distributions of ACC and F1 values of 50 different testing sets based on the RBF-based SVM model with 61 optimal immune signatures retained at least 80 times.

Figure 2

Construction and evaluation of the risk prognostic model based on immune features for GBM patients. (A) Forest plot summary of univariate Cox regression analysis of immune features significantly associated with overall survival. (B) Kaplan–Meier survival analysis of GBM patients that were divided into high-, medium-, and low-risk groups using a cutoff determined by X-tile. (C–E) ROC validation of the prognostic value of the predictive signature for predicting 1-, 3-, and 5-year survival of GBM patients, respectively. (F) Forest plot summary of the univariable analysis of IDH status, MGMT status, gender, age, and risk score. (G) Forest plot summary of the multivariable analysis of risk score, IDH status, MGMT status, and age. Here, the wildtype and mutant status of IDH and the unmethylated and methylated status of MGMT promoter were both converted to 1 and 0 respectively.

Figure 3

Landscapes of tumor immune microenvironment and clinicopathological characteristics of three GBM subtypes. (A) Heatmap depicting the association between GBM subtypes and immune cell infiltration. (B) The Kaplan–Meier curve for the OS of 397 GBM patients in three GBM subgroups. (C–E) Box plots for exploring the differences of CYT, ICG, APM, TIS, adaptive immune, innate immune, immune and stromal scores, and tumor purity among GBM subtypes. (F) Differences in the expressions of T-cell exhaustion markers between the three GBM subtypes. (G, H) Differences in the expressions of glioma antigens across the three GBM subtypes. (I, J) Different proportions of various immune and stromal cells in the GBM subgroups. For all box plots, the Kruskal–Wallis test was used to determine the significance of differences among GBM subtypes, and p-values are shown on the top of each box plot. ^* p< 0.05; ^** p< 0.01; ^*** p< 0.001; ^**** p< 0.0001; ns, no significant difference.

Figure 4

Analysis of the differences in the enrichment scores of KEGG pathways, Reactome categories, and GO terms demonstrated by GSVA among GBM subtypes. (A) Heatmap describing the top 10 significantly differential signatures, including KEGG pathways, Reactome, and GO terms between MES and CL. (B) Variants in KEGG pathways, Reactome categories, and GO terms between MES and PN.

Figure 5

(A) Comparisons of GET scores among the three GBM subtypes. (B–G) The correlations between GET score and the APM, CYT, innate immune, stromal, TIS scores, and tumor purity, respectively, in MES patients by Pearson’s correlation analysis. (H) Immunotherapy response prediction by SubMap analysis indicates a significant difference in anti-PD1 therapy response across the GBM subtypes.

Figure 6

Drug sensitivity evaluation based on hub genes and potential drug prediction for patients within the MES subtype. (A) The bubble plot showed the correlation between the mRNA expression of genes that were upregulated in the mesenchymal subtype and GDSC drug sensitives. (B–D) Structures of the three most significant bioactive chemicals sharing common MOA of MEK inhibitor by CMap analysis.