Machine learning revealed stemness features and a novel stemness-based classification with appealing implications in discriminating the prognosis, immunotherapy and temozolomide responses of 906 glioblastoma patients

• Published in: Briefings in bioinformatics Vol. 22; no. 5
• Main Authors: Wang, Zihao, Wang, Yaning, Yang, Tianrui, Xing, Hao, Wang, Yuekun, Gao, Lu, Guo, Xiaopeng, Xing, Bing, Wang, Yu, Ma, Wenbin
• Format: Journal Article
• Language: English
• Published: England Oxford University Press 02.09.2021
• Subjects: Problem Solving Protocol; stemness subgroup; integrated multiomic analysis; glioblastoma; mRNAsi; immunotherapy
• DOI: 10.1093/bib/bbab032
• ISSN: 1467-5463; 1477-4054

Glioblastoma (GBM) is the most malignant and lethal intracranial tumor, with extremely limited treatment options. Immunotherapy has been widely studied in GBM, but none can significantly prolong the overall survival (OS) of patients without selection. Considering that GBM cancer stem cells (CSCs) play a non-negligible role in tumorigenesis and chemoradiotherapy resistance, we proposed a novel stemness-based classification of GBM and screened out certain population more responsive to immunotherapy. The one-class logistic regression algorithm was used to calculate the stemness index (mRNAsi) of 518 GBM patients from The Cancer Genome Atlas (TCGA) database based on transcriptomics of GBM and pluripotent stem cells. Based on their stemness signature, GBM patients were divided into two subtypes via consensus clustering, and patients in Stemness Subtype I presented significantly better OS but poorer progression-free survival than Stemness Subtype II. Genomic variations revealed patients in Stemness Subtype I had higher somatic mutation loads and copy number alteration burdens. Additionally, two stemness subtypes had distinct tumor immune microenvironment patterns. Tumor Immune Dysfunction and Exclusion and subclass mapping analysis further demonstrated patients in Stemness Subtype I were more likely to respond to immunotherapy, especially anti-PD1 treatment. The pRRophetic algorithm also indicated patients in Stemness Subtype I were more resistant to temozolomide therapy. Finally, multiple machine learning algorithms were used to develop a 7-gene Stemness Subtype Predictor, which were further validated in two external independent GBM cohorts. This novel stemness-based classification could provide a promising prognostic predictor for GBM and may guide physicians in selecting potential responders for preferential use of immunotherapy.