GBM Bulk RNA-seq Deconvolution & Survival Analysis

A reproducible, multi-cohort framework for characterizing the association between myeloid infiltration, mesenchymal transition, and clinical outcome in glioblastoma (GBM, IDH-wildtype) using bulk RNA-seq deconvolution and survival analysis.

Scientific premise

GBM has a myeloid-dominated, immunosuppressive tumor microenvironment, and myeloid cells actively drive proneural→mesenchymal transition of malignant cells. In bulk RNA-seq, the "mesenchymal" signal conflates malignant-cell state with microenvironment infiltrate. This project uses single-cell-reference deconvolution to separate malignant-MES from myeloid signal, then tests their independent and joint association with overall survival in adjusted multivariable Cox models, with discovery → independent validation across cohorts. See docs/gap_analysis.md.

Directory layout

gbm-bulk-deconvolution/
├── data/
│   ├── raw/                # downloaded matrices + clinical (per cohort)
│   └── preprocessed/       # harmonized TPM/counts + reference + QC
├── scripts/                # numbered pipeline scripts (see scripts/README.md)
├── results/
│   ├── figures/
│   │   ├── main/           # main-text figures
│   │   └── supplement/     # supplementary figures
│   └── tables/
│       ├── main/           # main-text tables (incl. dataset catalog)
│       └── supplement/     # supplementary tables (incl. evidence table)
├── docs/                   # per-step documentation + review + blueprint
│   ├── literature_review.md
│   ├── gap_analysis.md
│   ├── dataset_catalog.md
│   ├── analysis_blueprint.md
│   └── 01_download.md ... 08_figures_tables.md
├── README.md               # this file
└── environment.yml         # reproducibility (R package versions)

Quickstart

Run the full pipeline locally (one command)

The repository bundles all required raw data (TCGA, CGGA, GLASS, Ivy GAP, signatures, and the pre-built single-cell reference), so the entire pipeline runs end-to-end from this folder.

# 1. Install R (>= 4.2) and Python (>= 3.10), then:

# 2. Restore the pinned R package library (into ./.Rlib; ~289 packages)
Rscript scripts/setup_env.R

# 3. Install the Python dependencies (GLASS cohort-build steps)
pip install -r requirements.txt

# 4. Run the whole pipeline (raw data -> publication figures)
Rscript run_all.R
#   or:  ./run_all.sh

run_all.R runs every step in dependency order and skips any step whose output already exists, so it is restartable and cheap to re-run. Useful modes:

Rscript run_all.R --smoke   # fast check: rebuild publication figures from
                            #   existing tables only (no heavy compute)
Rscript run_all.R --force   # re-run every step, ignoring existing outputs

A running log is written to results/tmp/run_all.log.

Explore the project

1. Review the science: docs/literature_review.md, docs/gap_analysis.md.
2. Review the data plan: docs/dataset_catalog.md and results/tables/main/gbm_rnaseq_datasets_VALIDATED.csv.
3. Review the methods: docs/analysis_blueprint.md.
4. Individual steps live in scripts/ and can be run standalone with Rscript scripts/<step>.R (each sources scripts/00_config.R for paths).

Cohorts

• Discovery: TCGA-GBM (GDC STAR-Counts).
• Validation: CGGA mRNAseq_693 and mRNAseq_325.
• Longitudinal: GLASS (Synapse).
• Spatial/context: Ivy GAP + selected GEO immunotherapy cohorts.

Methods at a glance

• Deconvolution: BayesPrism + DWLS (optional MuSiC/Bisque) with a GBM-specific single-cell reference (malignant states + microglia/monocyte-derived macrophages/ MDSCs/DCs/lymphocytes). Cross-method concordance reported.
• MES scoring: Neftel MES-like meta-module on the malignant compartment.
• Survival: Kaplan-Meier + adjusted multivariable Cox (age, MGMT, IDH, treatment), Schoenfeld PH diagnostics, EPV ≥ 10, C-index with overfit caveat, BH-FDR.
• Validation: direction-concordant replication in ≥ 1 independent cohort.

Reproducibility

R (>= 4.2) and Python (>= 3.10). R package versions are pinned exactly in renv.lock (289 packages, R 4.4.3, Bioconductor 3.20) and restored into the project-local .Rlib/ by scripts/setup_env.R; Python dependencies are pinned in requirements.txt. All file paths are resolved centrally by scripts/00_config.R (override the project root with the GBM_PROJ environment variable).

Bundled data. All required inputs are committed under data/ so the pipeline runs offline:

• data/raw/GLASS/ — GLASS TPM matrix + clinical (requires free Synapse registration to re-download; bundled here so no login is needed).
• data/raw/IvyGAP/ — Ivy GAP FPKM + region annotations.
• data/raw/signatures/ — Neftel meta-modules, Verhaak centroids, gene sets.
• data/preprocessed/reference/gbm_sc_reference.rds — the GBmap Core single-cell reference (22,719 genes x 4,218 cells, 16 cell types, 95 donors; Ruiz-Moreno et al., Nat Commun 2023). scripts/03_build_reference.R validates this file; an optional rebuild from the full 8 GB atlas is available via GBM_REBUILD_REFERENCE=1 (see that script's header).

TCGA (GDC) and CGGA are re-downloadable anonymously by scripts/01_download_data.R; their raw matrices are also bundled so the download step is skipped on re-run. The Pombo Antunes TAM-BDM marker set is a curated gene list (used in scripts/13_tam_bdm_markers.R), not a raw matrix, so no large Pombo files are bundled.

License / data-use

TCGA, CGGA, GLASS, Ivy GAP, and GEO data are subject to their respective terms of use; GLASS requires free Synapse registration and terms acceptance. Deconvolution uses commercially-licensed methods only (BayesPrism, DWLS, MuSiC, Bisque); CIBERSORTx/EPIC are intentionally excluded.