Machine-learning-assisted integration of single-cell and bulk transcriptomics characterizes glioblastoma ecosystem states and immune-associated programs

Glioblastoma (GBM) is characterized by marked malignant-cell plasticity and a complex tumor ecosystem composed of heterogeneous malignant and microenvironmental cell populations. We integrated single-cell and bulk transcriptomic data to characterize reproducible GBM cell and ecosystem-associated states, evaluate cross-platform concordance, and examine immune-associated transcriptional programs while distinguishing descriptive associations from mechanistic inference.

The deposited Smart-seq2 (7,930 cells) and 10x (16,201 cells) expression matrices from GSE131928/SCP393 were analyzed on their verified expression scales. Official annotations and malignant-state modules were audited, four broad malignant states were recomputed, and a tumor-held-out machine-learning label-transfer framework was applied to the 10x dataset. Public TCGA-GBM RNA-seq data were used to evaluate state- and microenvironment-associated transcriptional programs and to perform nested prediction benchmarking. Selected ecosystem-associated genes were additionally examined by RT-qPCR in glioma cell lines and normal human astrocytes.

Among Smart-seq2 cells, 6,576 malignant cells had valid state scores, whereas 287 alias-affected records were excluded from state-score denominators. The machine-learning label-transfer model achieved an out-of-fold accuracy of 0.998, and malignant-state fractions showed moderate-to-strong concordance across six shared tumor labels. At the tumor-label level, MES-like fraction was not significantly associated with macrophage fraction (Spearman rho = 0.171, P = 0.395), indicating that ecosystem-level transcriptional association should not be interpreted as a simple directional immune-to-state mechanism. In TCGA-GBM, MES and myeloid-associated programs correlated with immune and stromal expression signatures, whereas survival and 12-month prediction performance remained modest and specification-sensitive. RT-qPCR further demonstrated increased STAT3, CHI3L1, SPP1, and VEGFA expression in glioma cell lines relative to normal human astrocytes.

Integrated single-cell and bulk transcriptomic analysis supports reproducible characterization of malignant-cell heterogeneity and immune-associated ecosystem features in IDH-wildtype GBM. Machine-learning-assisted cross-platform analysis provides a reproducible framework for evaluating tumor ecosystem-associated states, while the observed immune-mesenchymal associations remain descriptive rather than causal. Experimental RT-qPCR provides expression-level support for selected ecosystem-associated genes and highlights testable hypotheses for future functional and spatial studies.