Analysis of Transcription Factor Activity in PBMC Dataset

Initial Project Plan

The primary goal was to preprocess a public human peripheral blood mononuclear cell (PBMC) dataset and infer the activity of key transcription factors (TFs) regulating cellular identity using an AUCell-like approach, followed by visualization. The planned workflow included:

• Setup: Install necessary bioinformatics libraries.
• Data Preparation: Load and preprocess raw single-cell data.
• Load Knowledge Base: Import regulatory databases (e.g., DoRothEA) mapping TFs to target genes.
• Network Building: Construct gene regulatory networks (regulons).
• Activity Calculation: Score regulon activity in each cell (AUCell).
• Integration: Merge AUCell scores into the main AnnData object.
• Visualization: Generate plots to visualize regulon activity.
• Completion: Finalize the analysis.

Execution and Troubleshooting

Initial data preparation was smooth, but challenges with bioinformatics libraries necessitated a pivot to a manual AUCell-like implementation.

Phase 1: Data Loading and Preprocessing

All standard preprocessing steps were executed using the scanpy library on the pbmc3k dataset.

• Setup and Loading: Installed libraries and loaded the dataset into an AnnData object.
  python SNIPPET

  View Source Code

  Click to expand interactive code modal

  import scanpy as sc
  adata = sc.datasets.pbmc3k()

• Quality Control (QC): Calculated, visualized, and filtered data based on QC metrics to remove low-quality cells and genes.
  python SNIPPET

  View Source Code

  Click to expand interactive code modal

  sc.pp.filter_cells(adata, min_genes=200)
  sc.pp.filter_genes(adata, min_cells=3)

• Normalization and Feature Selection: Normalized data, applied log-transform, and identified highly variable genes.
  python SNIPPET

  View Source Code

  Click to expand interactive code modal

  sc.pp.normalize_total(adata, target_sum=1e4)
  sc.pp.log1p(adata)
  sc.pp.highly_variable_genes(adata, min_mean=0.0125, max_mean=3, min_disp=0.5)

• Dimensionality Reduction and Clustering: Scaled data, performed PCA, computed a neighborhood graph, ran UMAP, and applied Leiden clustering.
  python SNIPPET

  View Source Code

  Click to expand interactive code modal

  sc.pp.scale(adata, max_value=10)
  sc.tl.pca(adata)
  sc.pp.neighbors(adata, n_neighbors=10, n_pcs=40)
  sc.tl.umap(adata)
  sc.tl.leiden(adata)

Key Outputs:

• QC Violin Plots: figures/violinqc_violin_plots.png
• Highly Variable Genes Plot: figures/filter_genes_dispersionhighly_variable_genes.png
• PCA Variance Ratio Plot: figures/pca_variance_ratiopca_variance_ratio.png
• UMAP of Leiden Clusters: figures/umap_leiden_clusters.png

Phase 2: Overcoming Library Challenges

Significant roadblocks were encountered during the core analysis phase.

• Challenges with decoupler: Persistent AttributeError and ModuleNotFoundError issues indicated installation or compatibility problems, unresolved despite reinstallation.
• Challenges with pySCENIC: Attempts to use pySCENIC failed due to inaccessible scenicplus package and ModuleNotFoundError issues with core functions for AUCell score calculation.

Phase 3: Successful Manual Implementation

A manual AUCell-like scoring method was implemented to overcome library issues.

• Loading the Knowledge Base: Loaded and formatted DoRothEA TF-target gene relationships using omnipath and pandas, creating a clean regulons dictionary.
• Custom Scoring Script: Developed a Python script to calculate AUCell-like scores based on gene expression ranks and the regulons dictionary, stored in adata.obsm['aucell_estimate'].
• Final Visualization: Visualized the inferred activity of TF E2F8 on a UMAP plot, showing differential activity across clusters.
  • Output: figures/umap_E2F8_activity.png

Conclusion and Summary of Phases

Despite challenges with decoupler and pySCENIC, the project achieved its core goal through a manual AUCell-like implementation, successfully inferring and visualizing TF activity.

• Phase 1: Data Acquisition and Preprocessing: Complete. Successfully acquired, cleaned, normalized, and clustered the single-cell data.
• Phase 2: Core Analysis and Inference: Complete (with adaptation). Substituted the failed VIPER method (via decoupler) with a custom AUCell-like scoring method.
• Phase 3: Validation and Visualization: Partially Complete. Visualized TF activity as planned, but further statistical validation or model export was not pursued, as the primary goal was achieved.