CMU × NVIDIA Hackathon — January 7–9, 2026

OmniGenome

Federated Pangenomes & Genomic Background Modeling

---

Slide Deck

🧠 Problem & Motivation

Pangenomes offer a powerful alternative to single-reference genomes by representing genetic variation across many individuals. However, current pangenome graphs typically include only tens of samples, while modern biobanks (UK Biobank, All of Us, HPRC) now contain thousands of long-read genomes spanning diverse populations.

A major barrier remains:

There is no practical way for independent biobanks to collaboratively construct population-scale pangenomes without sharing raw genomic data.

This leads to:

• Silos across cohorts
• Persistent population bias
• Limited downstream interpretability

---

🎯 Project Vision

OmniGenome explores whether federated learning principles can be applied to pangenome construction and genetic association analysis.

Instead of sharing raw FASTA files, participating sites:

• Build local pangenome graphs
• Share graph representations only
• Aggregate information per chromosome

We demonstrate this idea through:

1. Federated pangenome graph construction (HPRC)
2. Genomic background hashing for phenotype association at the APOE locus

---

🚀 Quick Start (Demo)

This quick start runs a single local PGGB job to demonstrate the core building block used throughout the project.

Prerequisites

• Docker (with Compose)
• ~8 GB RAM
• bgzip-compatible FASTA input

Run Example

# Start Docker
systemctl start docker

# Build containers
docker compose build

# Run PGGB on an example FASTA
docker compose run pggb pggb \
  -i /data/input.fa.gz \
  -o /output/my_graph \
  -n 12 \
  -t 8 -p 90 -s 10000

⚠️ FASTA files must be bgzip-compressed, not gzip.

This produces a local GFA pangenome graph, which is the unit shared in the federated workflow.

---

🧬 Project Overview

Two Complementary Components

1️⃣ Federated Pangenome Construction

Simulates how multiple biobanks can collaboratively build pangenomes while keeping raw genomic data local.

2️⃣ Genomic Background Hashing (APOE Case Study)

Uses localized pangenome context to encode anonymized haploblock structure for federated association analysis.

Together, these form an end-to-end framework for privacy-preserving, population-scale pangenomics.

---

🧬 High-Level Architecture

---

🔬 Component 1: Federated Pangenome Graph Construction

Data

• Human Pangenome Reference Consortium (HPRC)
• Chromosomes 19 and 22 (tractable proof-of-principle)

Core Idea

Each simulated “site”:

• Holds private FASTA data
• Builds a local pangenome graph using PGGB
• Shares only the resulting GFA graph

Graphs are then:

• Aggregated per chromosome
• Used to refine local graphs iteratively

⚠️ Graphs are never combined across chromosomes

---

Minimal Pipeline Summary

1. Download HPRC assemblies
2. Extract chr19 / chr22
3. Subsample individuals to simulate biobank cohorts
4. Run PGGB locally
5. Aggregate graphs with vg combine
6. (Optional) Feedback using minigraph

---

Example: Local Graph Construction

docker compose run pggb pggb \
  -i /data/input.fa.gz \
  -o /output/run_name \
  -n 20 \
  -t 8 -p 90 -s 10000

---

🧬 Component 2: Genomic Background Hashing (APOE)

To demonstrate downstream utility, we focus on the APOE locus, a major genetic risk factor for Alzheimer’s disease.

Motivation

GWAS hits often lack genomic context:

• Identical risk alleles can appear in different haplotypic backgrounds
• These backgrounds may influence penetrance or downstream effects

Pangenomes provide a natural representation of this structure.

---

Workflow

1. Identify APOE-associated loci from GWAS summary statistics
2. Extract the corresponding pangenome subgraph
3. Encode anonymized haploblock structure as hashes
4. Use hashes as a federated genomic background feature

This enables:

• Privacy-preserving association testing
• Context-aware interpretation of risk alleles
• Cross-cohort comparison without raw genotype sharing

---

🧰 Key Tools

Tool	Role
PGGB	Local pangenome construction
vg	Graph conversion & aggregation
minigraph	Incremental graph feedback
odgi	Graph slicing & analysis
Docker	Reproducibility

---

👥 Team

CMU × NVIDIA Hackathon 2026

• Rob Loughnan
• Adam Kehl
• Jedrzej Kubica
• Kumar Koushik Telaprolu
• Jeff Winchell
• Sanjnaa Sridhar
• Dhruv Gor
• Samarpan Mohanty

---

Thank you to the organizers of the hackathon!

📚 References

• Wightman DP et al. Nature Genetics (2021)
• Garrison E et al. Nature Methods (2024)
• Garrison E et al. Nature Biotechnology (2018)
• Guarracino A et al. Bioinformatics (2022)
• Sirén J et al. Science (2021)
• Liao W-W et al. Nature (2023)
• Li H et al. Genome Biology (2020)

---

📄 License

MIT License