Stability-Aware Pathway-Level Unsupervised Learning Recovers Interpretable Metabolic Structure in Glioblastoma

Clusters from the high-dimensional transcriptomic space may be difficult to interpret and may vary from one cohort to another and one platform to another. In previous GBM studies, different transcriptomic GBM subtypes were identified and metabolic GBM states, such as glycolysis/mitochondrial and mesenchymal/proneural, were defined [1,2]. We thus pose the question of whether it is possible to recreate coherent metabolic structure without directly clustering on thousands of genes when pathway representations are biologically structured and consensus is learned using perturbations. For this analysis, we used 573 GBM samples with primary IDH status that were sequenced as RNA-seq and/or hybridized to microarrays from TCGA-GBM, CGGA-693 and CGGA-325. Following gene harmonization, TMM normalization, voom transformation and ComBat correction, 14,123 common genes were kept. Using ssGSEA, expression was normalized to pathway-level activity for 20 curated Hallmark/KEGG metabolic gene sets, and the 10 pathways with the largest SD of normalized expression were selected as learning features. For each k (k=2–8), we subsampled 80% of tumors 1,000 times for each k, calculated co-clustering similarities and consensus matrices, and chose k by consensus CDF and delta-area. The delta-area curve showed a marked decrease from k=2 to k=3, supporting the parsimonious k=2 model. The resulting partition was nearly balanced (n=286, HG-GBM; n=287, OM-GBM). Hypoxia, glycolysis, reactive-oxygen-species, glutathione and stress-associated activities were higher in HG-GBM, whereas oxidative-phosphorylation, TCA-cycle, fatty-acid and amino-acid metabolic activities were higher in OM-GBM. The learned separation was validated beyond the clustering features, with 503 differentially expressed genes distinguishing the states. Hallmark GSEA showed strong enrichment of inflammatory, hypoxia-related, EMT and TNFα/NFκB-associated programs in HG-GBM (FDR ≈ 4×10−10), and neuronal/synaptic programs in OM-GBM. Our contribution is methodological rather than a claim of new GBM biology. We demonstrate that pathway-level compression combined with perturbation-based consensus learning can preserve biologically meaningful latent states in heterogeneous transcriptomic cohorts, with convergence to established GBM axes serving as a reproducibility signal [1,2]. Three additional cohorts are reserved for external validation of cross-cohort transferability (GSE16011, REMBRANDT, GSE4290). This provides a concise framework for stability-aware unsupervised learning in biomedical transcriptomics.