Acute Myeloid Leukemia (AML) lives in the bone marrow niche. Leukemic cells reshape that niche so stromal, endothelial and immune cells help protect them from therapy.
Gfi1 is a transcriptional repressor. When it is lost or down-regulated:
- Target genes are no longer repressed
- Those genes produce secreted signals (cytokines, chemokines, matrix factors)
- The signals bind receptors on neighboring niche cells
- The niche becomes a protective shield → worse outcome
Question this repo answers:
Which secreted signals rise when Gfi1 is low, which of them associate with clinical risk, and which receptors on niche cells do they engage?
The pipeline maps this axis using public real data (TCGA-LAML + GSE116256) and produces tables, figures and a summary report.
Gfi1 low / mutant AML cell
│
▼ stops repressing targets
Secreted ligands (TGFB1, S100A8/A9, CXCL8, VEGFA, IL10, …)
│
▼ leave the cell
Bind receptors on progenitors / monocytes / dendritic / T cells
│
▼
Protective niche → higher risk, therapy resistance
| Dataset | What it is | How it was used |
|---|---|---|
| TCGA-LAML (UCSC Xena HiSeqV2) | 173 AML patients, bulk RNA-seq + survival | Machine-learning ranking of secreted genes vs high-risk status |
| GSE116256 (van Galen et al., Cell 2019) | AML diagnosis + healthy bone marrow scRNA-seq | GFI1-stratified differential expression + ligand–receptor scoring |
An XGBoost model was trained on expression of candidate secreted genes to predict high-risk status (death with short overall survival).
| Metric | Value |
|---|---|
| Test-set AUC | 0.685 |
| 3-fold CV AUC | 0.624 ± 0.093 |
Top genes by predictive importance
| Rank | Gene | Importance | Role |
|---|---|---|---|
| 1 | IL10 | 0.069 | Immunosuppressive cytokine |
| 2 | PDGFB | 0.068 | Growth factor, stromal remodeling |
| 3 | CXCL10 | 0.057 | Inflammatory chemokine |
| 4 | HGF | 0.053 | Scatter factor / niche support |
| 5 | S100A8 | 0.043 | Inflammatory alarmin |
| 6 | CXCL2 | 0.040 | Neutrophil recruitment |
| 7 | CXCL1 | 0.038 | Inflammatory chemokine |
| 8 | TGFB1 | 0.032 | Quiescence, immune modulation |
| 9 | LGALS1 | 0.032 | Galectin-1, immune evasion |
| 10 | ANGPT1 | 0.029 | Vessel / niche stability |
Figure: Secreted genes ranked by how strongly they help predict high-risk status in TCGA-LAML.
AML cells were split by GFI1 expression (low vs high). Secreted genes were ranked by log2 fold-change (GFI1-Low / GFI1-High).
Note: GFI1 is sparsely detected in this Seq-Well dataset (~33 cells with clear expression in the subsample), so formal FDR is limited. Rankings are hypothesis-generating.
| Gene | log2FC trend | Biological note |
|---|---|---|
| LGALS3 | ↑ in GFI1-low | Galectin-3, adhesion / survival |
| CXCL8 | ↑ | Neutrophil / inflammatory signaling |
| TIMP1 | ↑ | Matrix remodeling |
| CCL5 | ↑ | Chemokine, immune recruitment |
| TGFB1 | ↑ | Quiescence, niche remodeling |
| PDGFA | ↑ | Stromal growth factor |
| TNF | ↑ | Inflammatory cytokine |
| VEGFA | ↑ | Angiogenesis |
| SPP1 | ↑ | Osteopontin, adhesion / resistance |
| ANGPT2 | ↑ | Vessel destabilization |
Figure: Secreted genes ordered by log2FC between GFI1-Low and GFI1-High AML cells (GSE116256).
Figure: GFI1 expression distribution and cell-type composition in the analyzed GSE116256 subset.
Ligand–receptor scores were computed from AML GFI1-low cells toward niche populations (progenitors, monocytes, dendritic cells, T cells, erythroid).
41 interactions were scored. The strongest axes:
| Ligand | Receptor | Receiver cell type | Interpretation |
|---|---|---|---|
| S100A9 | CD44 | Progenitor, Monocyte, Dendritic | Inflammatory alarmin → adhesion / survival |
| S100A8 | CD44 | Progenitor, Monocyte, Dendritic | Same axis |
| TGFB1 | TGFBR1 / TGFBR2 | Erythroid, Dendritic, T cell | Quiescence / immune modulation |
Figure: Ligand–receptor interaction scores (AML GFI1-low → niche cell types).
Across bulk clinical data and single-cell niche data, the same families of signals keep appearing:
- Inflammatory alarmins (S100A8/A9) engaging CD44
- TGFB1 signaling into niche and immune cells
- Chemokines / growth factors (CXCL family, HGF, PDGF, VEGFA, IL10)
These are coherent with a model in which low Gfi1 activity favors a secreted program that remodels the bone-marrow microenvironment and associates with higher clinical risk.
Next steps suggested by this map:
- Validate top axes in larger scRNA-seq atlases (BoneMarrowMap, AML scAtlas)
- Structural modeling (AlphaFold3) of priority pairs (e.g. S100A8/A9–CD44, TGFB1–TGFBR)
- Virtual screening / experimental disruption of those interfaces
See docs/structural_followup.md for a concrete protocol.
cd Gfi1_AML_Signaling_Mapping
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
# Demo (synthetic biology, full end-to-end)
python -m src.run_pipeline
# Real data (requires TCGA + GSE116256 files under data/raw/)
# See docs/data_sources.md for download links
python -m src.real_data_pipeline # or the streaming scripts used for this reportGfi1_AML_Signaling_Mapping/
├── README.md ← this file
├── LICENSE
├── requirements.txt
├── src/ ← analysis code
│ ├── run_pipeline.py ← demo end-to-end
│ ├── real_data_pipeline.py ← real-data entry point
│ ├── data_generation.py
│ ├── differential_expression.py
│ ├── ml_prioritization.py
│ ├── cell_communication.py
│ ├── visualize.py
│ └── report.py
├── docs/
│ ├── methods.md
│ ├── data_sources.md
│ └── structural_followup.md
├── results/
│ ├── SUMMARY_REPORT_REAL.md
│ ├── tables/ ← CSV results (real + demo)
│ └── figures/ ← PNG figures used above
└── data/
├── raw/ ← place TCGA / GEO downloads here
└── processed/
| File | Content |
|---|---|
results/SUMMARY_REPORT_REAL.md |
Full narrative report from real data |
results/tables/ml_feature_importance_real.csv |
TCGA gene ranking |
results/tables/ml_performance_real.csv |
AUC metrics |
results/tables/deg_gfi1_low_vs_high_real.csv |
Full DEG table (GSE116256) |
results/tables/secreted_candidates_real.csv |
Secreted genes ranked by log2FC |
results/tables/lr_interaction_scores_real.csv |
Ligand–receptor scores |
results/figures/*.png |
All figures embedded in this README |
Cite the original data sources:
- van Galen et al., Cell 2019 (GSE116256)
- TCGA LAML / UCSC Xena
- This repository for the analysis workflow
Released under the MIT License — see LICENSE.
This project turns a clear biological hypothesis about Gfi1 and the AML niche into a transparent map built on public clinical and single-cell data.