TRM_project

Single-cell RNA-seq analysis of tissue-resident memory T cells (TRM) in psoriatic skin, combining Scanpy preprocessing, pySCENIC gene regulatory network inference, and basin stability modeling to identify transcription factors that maintain the IL-22-high TRM attractor state.

Data Source

• GEO Accession: GSE204765
• Samples: BCH01 non-lesional (NL) vs lesional (L) skin biopsies
• Format: 10x scRNA-seq (h5ad)

Project Structure

TRM_project/
├── raw_data/                  GEO h5ad files (not tracked by git)
├── processed_data/            T cell / TRM subsets, expression matrix, metadata
├── scripts/
│   ├── trm_scanpy.py          Scanpy pipeline: QC → T cell gating → TRM gating → UMAP
│   ├── setup_pyscenic.sh      Conda environment setup for pySCENIC
│   ├── IL22TRMPOU2F1.ipynb    Main analysis: basin stability & TF perturbation
│   ├── locking_tf_rank.ipynb  Delta AUC ranking of locking TFs
│   └── depth_analysis_simulation.ipynb  Depth-based analysis & power simulations
├── GRN_SCENIC/                Regulons, GRN, AUC matrix, motif/ranking databases
├── figures_results/           Figures (pdf/png/svg), tables, and result CSVs
└── manuscript/                Cell Systems submission materials

Analysis Pipeline

1. Preprocessing (trm_scanpy.py)

   • Load NL and L h5ad from GEO
   • QC and module scoring (T cell, CD4, CD8, TRM core, circulating)
   • Gate T cells (TRAC+ or score > 0.2), then gate TRM (TRM-core high, circulating low)
   • Concatenate, HVG selection, PCA, UMAP, Leiden clustering
   • Differential expression: L vs NL (Wilcoxon)

2. GRN Inference (pySCENIC)

   • Regulon detection with SCENIC motif databases (hg38, v9)
   • AUC scoring per cell (trm_auc.csv)

3. Basin Stability Modeling (IL22TRMPOU2F1.ipynb)

   • Define IL-22-high attractor state from TRM metadata
   • Fit logistic regression classifier in regulon AUC space
   • Compute basin stability curves under noise perturbation
   • Single-TF leave-one-out: identify TFs most critical for stability
   • Pairwise TF removal: identify cooperative TF pairs (e.g., POU2F1 + CHD1/RBBP5)
   • Permutation test for synergy significance

4. Depth-Based Reanalysis (depth_analysis_simulation.ipynb)

   • Cross-validated attractor depth estimation
   • Power simulations for cooperative TF detection

Key Findings

• POU2F1 is a central locking TF for the IL-22-high TRM state
• Cooperative TF pairs (e.g., CHD1 + RBBP5) show synergistic effects on basin stability
• TRM in lesional skin exhibit distinct regulon activity compared to non-lesional skin

Requirements

scanpy
anndata
pyscenic
scikit-learn
pandas
numpy
matplotlib
seaborn

Setup pySCENIC environment:

bash scripts/setup_pyscenic.sh

Large Files (not tracked)

The following are excluded via .gitignore and must be obtained separately:

File	Source
raw_data/GSM6190732_BCH01-NL1.h5ad	GEO
raw_data/GSM6190734_BCH01-L1.h5ad	GEO
GRN_SCENIC/hg38__refseq-r80__*_rankings.feather	SCENIC databases
GRN_SCENIC/motifs-v9-nr.hgnc-*.tbl	SCENIC databases