DecodeME AlphaGenome Analysis Pipeline

A robust, modular pipeline for post-GWAS functional characterization using AlphaGenome. This tool takes summary statistics (REGENIE format), identifies significant loci, refines them using LD information and Bayesian fine-mapping, and predicts functional effects (RNA expression, chromatin accessibility, splicing) in relevant tissues.

Features

• Robust Clumping: Identifies independent genomic loci using distance-based clumping with deterministic sorting.
• Informed Selection: Pre-calculates the 95% Credible Set size for every locus, allowing you to prioritize high-value targets before running expensive AlphaGenome queries.
• LD-Informed Fine-Mapping: Integrates with LDlink API to filter variants based on genetic linkage ($R^2 \ge 0.1$) rather than just physical proximity.
• Two-Tier Fallback: If LD data is unavailable (e.g., rare variants missing from 1000G), the pipeline automatically:
  1. Attempts to use nearby "Neighbor Proxies".
  2. Falls back to a reduced physical window (±50kb) to focus on the immediate vicinity.
• QC Compliance: Respects author-recommended QC filtering by default, ensuring results are scientifically sound.
• Multimodal Scoring: Queries AlphaGenome for RNA-Seq, ATAC-Seq, and Splicing predictions.
• Tissue Specificity: Automatically filters results for:
  • Brain (UBERON:0000955)
  • Blood (UBERON:0000178)
  • T-Cells (CL:0000084)
• Performance: Uses awk for high-speed file scanning and implements aggressive caching for API calls and regional data.

Setup

Prerequisites

• Python 3.8+
• Dependencies: pandas, numpy, requests, tqdm, scipy
• AlphaGenome Python library installed and authenticated.

Environment Variables

You must set the following API keys in your environment:

# Required for functional annotation
export ALPHAGENOME_API_KEY="your_alphagenome_key"

# Optional (highly recommended) for LD-informed fine-mapping
export LDLINK_API_TOKEN="your_ldlink_token"

Usage

Basic Run

Run the pipeline with default settings (interactive mode). It will scan the GWAS file, show identified loci with their Credible Set sizes, and ask you which to analyze.

python3 analyze_decodeme.py

Advanced Options: Skipping QC

The pipeline defaults to strict author-recommended QC filtering. If you want to analyze variants that were excluded (e.g. for exploratory analysis of missing peaks like Chromosome 22), use the --skip-qc flag.

python3 analyze_decodeme.py --chrom 22 --skip-qc

Targeted Analysis

Run analysis on a specific chromosome or region to save time.

# Analyze only Chromosome 6
python3 analyze_decodeme.py --chrom 6

# Analyze a specific region
python3 analyze_decodeme.py --chrom 6 --start 25000000 --end 35000000

Custom Thresholds & Configuration

python3 analyze_decodeme.py 
  --sig-threshold 8.0 
  --window 1000000 
  --credible-set 0.99 
  --output my_custom_analysis

Automated / Batch Mode

python3 analyze_decodeme.py --non-interactive --no-ldlink

Pipeline Steps

1. Lead Identification: Scans the input summary statistics for variants exceeding the significance threshold.
2. Clumping: Groups significant variants into independent genomic loci.
3. Regional Extraction: For each locus, extracts all variants within the window.
4. Credible Set Pre-calculation: Defines a 95% Credible Set using Approximate Bayes Factors (ABF). The size of this set is displayed to help the user manage AlphaGenome workload.
5. LD Filtering (Tiered):
   • Primary: Queries LDlink for variants in LD ($R^2 \ge 0.1$) with the lead SNP.
   • Fallback 1: If lead is missing, tries nearby neighbors as proxies.
   • Fallback 2: If all fails, reduces the window to ±50kb to focus on the immediate vicinity.
6. Annotation: Queries AlphaGenome to predict functional impacts for every variant in the credible set.
7. Filtering & Save: Filters predictions for prioritized tissues and saves results to CSV.

Caching

The pipeline creates two cache directories:

• region_cache/: Stores extracted GWAS summary statistics for specific windows.
• ld_cache/: Stores LDlink API responses.
• *_leads.csv: Caches the initial scan of significant hits.