多樣性을 保存하는 多層 機械學習 프레임워크를 使用한 膠母細胞腫에서의 生存에 따른 바이오마커 發見 = - Survival-concordant biomarker discovery in glioblastoma using a diversity-preserving multi-tier machine-learning framework -

다국어 초록 (Multilingual Abstract)

Title : Survival-concordant biomarker discovery in glioblastoma using a diversity-preserving multi-tier machine-learning framework Background: Robust transcriptomic biomarker discovery in glioblastoma multiforme （GBM） is limited by cohort heterogeneity, small sample sizes, and algorithm-dependent feature selection. Conventional machine-learning pipelines often prioritize classification accuracy without considering whether gene-expression changes are concordant with survival risk. Therefore, we developed a survival-concordant, diversity-preserving multi-tier machine-learning framework for biomarker discovery across heterogeneous GBM transcriptomic cohorts. Results: RNA-seq data from heterogeneous public cohorts （TCGA, CGGA, and GEO） were integrated to evaluate three continuous disease-state transitions: Normal-to-Primary （NvP）, Normal-to-Recurrenct （NvR）, and Primary-to-Recurrenct （PvR）. The framework combined a 223-model first-tier machine-learning tournament, diversity-preserving ensemble selection, an expression–prognosis discordance filter, and bootstrap-LASSO-weighted secondary meta-ensemble scoring. The finalized locked-model pipeline achieved robust discriminative performance with AUC values of 0.993, 0.976, and 0.681 for the internal NvP, NvR, and PvR tracks, respectively. Projection onto independent external GEO cohorts （GSE121720 and GSE263588） demonstrated outstanding generalizability （AUCs of 0.966 and 1.000） alongside a 100% feature preservation rate. Integrated multi-cohort prioritization （TCGA and CGGA） successfully identified mechanistically distinct diagnostic modules led by SCN3B （NvP）, ORC1 （NvR）, and IGFBP2 （PvR）. Notably, IGFBP2 showed progressive, monotonic upregulation from normal tissue to primary and recurrent GBM, correlating with poorer survival. Single-cell and immune analyses revealed that the recurrent-specific IGFBP2 and collagen module （COL3A1, COL4A1） reflects malignant, vascular, and myeloid-associated microenvironmental remodeling. Conclusion: This study developed a survival-concordant, diversity-preserving multi-tier machine-learning framework for transcriptomic biomarker discovery across heterogeneous public cohorts. By integrating multi-cohort RNA-seq analysis, ensemble model selection, expression–prognosis discordance filtering, and bootstrap-LASSO-weighted secondary scoring, the framework prioritized biomarkers with both discriminatory value and survival-concordant directionality. Application to glioblastoma successfully decoded three mechanistically distinct disease-state programs: the loss of normal neural identity （SCN3B）, the acquisition of unrestrained proliferation （ORC1）, and extracellular matrix （ECM） remodeling during recurrence （IGFBP2 and the associated collagen module）. Notably, IGFBP2 emerged as a critical bridge marker across the disease continuum, reflecting malignant-cell, vascular, and immune-associated features within the recurrent microenvironment. These findings support the utility of the proposed framework for prioritizing biologically interpretable biomarkers from public transcriptomic datasets and provide a basis for future experimental validation and extension to other cancer types or multi-omic data. Key word: Glioblastoma, machine learning, multi-tier framework

목차 (Table of Contents)

• Ⅰ. Introductions 1
• Ⅱ. Meterials and Methods 4
• 1. Data acquisition and pre-processing 4
• 2. Differential expression analysis and feature-space construction 6
• 3. First-tier diversity-preserving machine-learning tournament 6
• 4. Expression-prognosis discordance filter (Gatekeeper) 8
• 5. Secondary machine-learning framework: Secondary ML meta-ensemble scoring 9
• 6. Training set validation 15
• 7. Survival, trajectory expression, and single-gene correlation GSEA 15
• 8. Biological insights of Markers 16
• 9. Immune microenvironment profiling 17
• 10. Test set validation 18
• 11. Statistical analysis and reproducibility 19
• Ⅲ. Results 20
• 1. Cohort integration and differential expression landscape 20
• 2. First-tier machine-learning tournament and diversity-preserving ensemble selection 23
• 3. Expression-prognosis discordance filtering 26
• 4. secondary meta-ensemble re-ranking 30
• 5. Training set validation and data leakage tests 33
• 1) Locked-model ROC-AUC analysis 33
• 2) Training set data leakage and robustness tests 33
• 6. Survival and trajectory expression analysis of top candidates 37
• 1) survival and trajectory expression analysis of markers 37
• 2) Single-gene correlation GSEA-GO/Reactome analysis 38
• 7. Biological insights of the finalized signatures 41
• 1) Protein network functional analysis 41
• 2) Transcription factor analysis 41
• 3) Drug repositioning analysis 42
• 4) Single-cell RNA-seq resolution analysis 42
• 8. Immune microenvironment profiling 47
• 9. External test set validation and clinical characterization · 52
• 1) Locked-model ROC-AUC analysis 52
• 2) Cross-cohort feature preservation and reproducibility 52
• 3) Survival and single-gene correlation GSEA of robust markers 53
• 4) Cellular context via single-cell RNA-seq validation 53
• Ⅳ. Discussion 58
• V. Conclusion 65
• Data availability 66
• Declaration of AI and AI-assisted technologies 67
• References 68
• Abstract 77