pcst-ad

Snakemake pipeline for identifying Alzheimer's disease-relevant protein modules by integrating three data modalities:

• Edges: Protein-protein interactions from STRING database
• Edge weights: Differential co-expression scores from FAVA (Control vs AD)
• Node weights: Gene-level p-values from MAGMA (AD GWAS)

The Prize-Collecting Steiner Tree (PCST) algorithm builds connected gene networks by balancing strong genetic signals against the cost of traversing weaker disease-specific co-expression links.

Pipeline Overview

graph LR
    A[counts] --> B[fava]
    B --> C[network]
    C --> D[scppin]
    D --> E[gsea]
    D --> F[heatmap]
    D --> G[plot]

Loading

Workflow Steps

• counts: Extract cell-type specific expression data filtered to MAGMA GWAS genes
• fava: Compute gene-gene co-expression scores for Control and AD conditions (GPU-accelerated)
• network: Compute differential correlation using Fisher z-transformation, filter edges to known STRING protein-protein interactions
• scppin: PCST module detection—edges weighted by differential co-expression, nodes weighted by GWAS significance
• gsea: Gene set enrichment analysis on detected modules
• heatmap: Cross-cell-type module comparison using Jaccard similarity
• plot: Visualize detected modules as network graphs with nodes colored by GWAS p-value significance

Input Data

Required input files in input/ directory:

• magma_gene_symbol_results.tsv - GWAS gene-level association results from MAGMA
• GSE174367_snRNA-seq_cell_meta.csv - Single-cell RNA-seq metadata with cell type and diagnosis annotations
• GSE174367_snRNA-seq_filtered_feature_bc_matrix.h5 - Expression count matrix in 10x Genomics format

Configuration

Create a config.yaml file:

BUCKET: gs://your-bucket

CONDITIONS:
  - Control
  - AD

CELLTYPES:
  - ODC
  - EX
  - INH
  - ASC
  - MG
  - OPC
  - PER.END

# GPU settings for fava rule
GPU_TYPE: nvidia-tesla-t4
GPU_MACHINE_TYPE: n1-highmem-2
GPU_BOOT_IMAGE: projects/rocky-linux-accelerator-cloud/global/images/family/rocky-linux-8-optimized-gcp-nvidia-latest

# GSEA gene sets
GENE_SETS:
  - Reactome_2022
  - DisGeNET

# STRING cutoff
STRING_CUTOFF: 500

Output Files

Per Cell Type (output/{celltype}/)

• {condition}_anndata.h5ad - Filtered AnnData object with highly variable genes
• {condition}_fava_pairs.tsv - Gene pair co-expression scores
• network.tsv - Differential correlation network filtered by STRING PPI
• scppin_object.pkl - PCST module detection results
• module_gsea_enrichment.tsv - Pathway enrichment results

Figures (figures/)

• {celltype}/{condition}_fava_histogram.png - Distribution of FAVA scores
• {celltype}/module_network.png - Network visualization of detected module with GWAS p-value coloring
• module_jaccard_heatmap.png - Cross-cell-type module similarity heatmap

Usage

Local execution

snakemake --use-conda

Google Cloud Platform execution

snakemake --profile profiles/gcp

Data Sources

GWAS Data:

• Bellenguez et al. (2022). "New insights into the genetic etiology of Alzheimer's disease and related dementias." Nature Genetics, 54, 412-436.
• Link: https://www.nature.com/articles/s41588-022-01024-z
• MAGMA gene-level results derived from this GWAS

Single-cell RNA-seq Data:

• Morabito et al. (2021). "Single-nucleus chromatin accessibility and transcriptomic characterization of Alzheimer's disease." Nature Genetics, 53, 1143-1155. (GEO: GSE174367)
• Link: https://www.nature.com/articles/s41588-021-00894-z

STRING Database:

• Protein-protein interaction network (v11.5)
• Link: https://string-db.org