Pseudobulk Function Prediction Pipeline

Author: Alex Adrian-Hamazaki

---

Overview

This repository contains the Nextflow pipelines and analysis notebooks for a thesis investigating how cellular composition of bulk RNA-seq data affects gene function prediction using the EGAD (Extending Gene set Analysis with Diagnostics) framework.

The core question: when bulk RNA-seq is used to build co-expression networks for gene function prediction (via neighbor voting / AUROC), how much of the predictive signal comes from the tissue's cell type composition rather than cross-cell co-variance?

---

Repository Structure

.
├── function_prediction_pipelines/   # Nextflow pipelines
│   ├── bulkEGADPipeline/            # EGAD on real GTEx bulk RNA-seq
│   ├── ctProfilePipeline/           # EGAD on scRNA-seq cell type profiles
│   ├── EGADSimulatedGOTerms/        # EGAD on simulated bulk with controlled composition
│   └── expansionSimulation/         # EGAD while varying number of cell type profiles
└── PaperFigures/                    # Notebooks generating all paper figures
    ├── Mainfigures/
    │   ├── SimulatedBulk/           # Figures from simulated bulk analyses
    │   ├── CellTypeProfileAnalysis/ # Figures from CT profile analyses
    │   └── MGC/                     # Multi-gene co-expression figures
    └── Supplemental/                # Supplemental figure notebooks

---

Methods Summary

Gene Function Prediction with EGAD

All pipelines use EGAD's neighbor voting algorithm:

1. Build a Pearson co-expression network from an expression matrix (samples × genes)
2. For each GO term, predict membership using network neighborhood labels
3. Report AUROC per GO term

Expression Data Sourcesn

Source	Type	Use
GTEx	Real bulk RNA-seq (multi-tissue)	Baseline function prediction performance
Brain scRNA-seq	Single-cell h5ad	Cell type profiles + bulk simulation ground truth

Function Labels

Pipelines

Pipeline	Input	Purpose
bulkEGADPipeline	GTEx h5ad	Function prediction on real bulk data
ctProfilePipeline	scRNA-seq h5ad	Function prediction from cell type profiles
EGADSimulatedGOTerms	scRNA-seq h5ad + composition JSON	Simulate bulk at controlled compositions; test real & simulated GO terms
expansionSimulation	scRNA-seq h5ad	Test how adding/removing cell type profiles changes function learnability

---

Dependencies

All pipelines are managed with Nextflow and Conda/Mamba environments.

• Nextflow ≥ 23.x
• Python (via conda env main_env): anndata, scanpy, pandas, numpy
• R: EGAD, tidyverse

---

Quickstart

Each pipeline has its own nextflow.config and bootstrap_pipe.sh. See the README in each subdirectory under function_prediction_pipelines/ for pipeline-specific instructions.

# Example: run the simulated GO terms pipeline
cd function_prediction_pipelines/EGADSimulatedGOTerms
nextflow run sim_pipe.nf -entry bootstrap

---

Analysis Notebooks

Paper figures are generated from Jupyter notebooks in PaperFigures/. See PaperFigures/README.md for a breakdown of what each notebook produces.