Assessing glioblastoma cell population stability through bootstrap resampling of scRNA-seq data

Abstract

Glioblastoma (GBM) exhibits extreme cellular heterogeneity, comprising diverse tumor cell states and non-malignant microenvironment populations. Single-cell RNA-sequencing (scRNA-seq) enables resolution of this complexity, yet a critical unmet challenge persists: cluster reproducibility in GBM scRNA-seq studies is rarely validated, and standard clustering algorithms may generate artifactual partitions indistinguishable from biologically meaningful populations. To address this gap, we propose a cluster-wise bootstrap stability framework integrated with explicit tumor–microenvironment separation, an approach not previously applied systematically to GBM scRNA-seq data. We analyzed a public dataset (GSE131928; 10 tumors, 15,072 cells after quality control) and identified 14 clusters annotated via marker gene validation. Bootstrap resampling (100 iterations) with Jaccard coefficient quantification revealed that non-malignant populations (microglia/macrophage, oligodendrocytes) exhibited the highest stability (Jaccard >0.97). Among tumor states, MES-AC transitional and MES-like clusters were most stable (Jaccard 0.99 and 0.82), whereas NPC-like, AC-like, and rare populations showed low stability (Jaccard <0.5). Stability correlated positively with marker gene specificity, within-cluster homogeneity, and silhouette scores. These results demonstrate that cluster-wise bootstrap assessment provides a practical, quantitative criterion for distinguishing robust from unreliable cell populations, supporting more confident biological interpretation and therapeutic target prioritization in GBM.