CMU x NVIDIA Hackathon, January 7-9, 2026

Pangenome and Genomic Cluster Modeling

Problem

Pangenomes have the potential to more accurately capture genomic diversity across humans, however current state-of-the-art graphs are only on tens of individuals. Large biobanks (UK Biobank, AllOfUS) have long read sequencing data on thousands of people that could be used in the service constructing a more diverse pangenome graph, however methods do not currently exists for pangenome graph construction that enables cross-talk between biobanks.

Intro and aim

For this project we are aiming to:

1. Perform federated pangenome graph construction (using HPRC data as a proof of principle)
2. Perform federated genomic background hashing for phenotype association of APOE locus (using 1k genomes data)

Contributors

How to use this repo

Methods

1) Federated pangenome graph construction

Flowchart

a) Download Data from HPRC

Download data from HPRC for generating graphs from:

download_path=/path/to/download/destination/
python ./HPRC_download_prep/download_hprc.py \
./HPRC_download_prep/assemblies.tsv \
$download_path

Extract Chromosome 19 and 22 from HPRC samples

As pangenome graph construction is very computationally intensive we will be running the process on chromosome19 and chromsome22 as a more tractable dataset.

Install Entrez Direct using conda

The contig names in the fasta files for HPRC are NCBI identifiers and need to be queried using edirect tools to convert these to convetional chromosme numbers (e.g. CM102454.1 -> chr22). This can be installed with conda using:

conda install bioconda::entrez-direct

From this, the following python script can be used to extract chr19 and chr22 from the assembly FASTA files for HPRC:

# This should be the path to the assemblies downloaded above
download_path='/path/to/download/destination/'
python ./HPRC_download_prep/make_hrcp_chr22_fasta.py \
/space/ceph/1/ABCD/users/rloughna/pangenome_construction/hprc_chr22_pansn_full.fa.gz \
--output-chr19 /space/ceph/1/ABCD/users/rloughna/pangenome_construction/hprc_chr19_pansn_full.fa.gz \
--n-individuals 20 \
--bgzip

b) Create Partitions of Data to Simulate Biobank Cohorts

WIP

c) Graph Construction

Prerequisites

• Docker CE with Compose plugin (v5.0+)
• Minimum 8GB RAM recommended

PGGB (Pangenome Graph Builder)

Builds pangenome graphs from multi-sample FASTA files.

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

VG (Variation Graph Toolkit)

Converts GFA graphs to Giraffe-compatible formats (GBZ, dist, min).

docker compose run vg autoindex --workflow giraffe \
  -g /data/<graph.gfa> \
  -p /data/<output_prefix>

Quick Start

# 1. Start Docker daemon
systemctl start docker

# 2. Build containers
docker compose build

# 3. Run PGGB to build graph
docker compose run pggb pggb \
  -i /data/input.fa.gz \
  -o /output/my_graph \
  -n 12 -t 8 -p 90 -s 10000

# 4. Convert to Giraffe format
docker compose run vg autoindex --workflow giraffe \
  -g /data/my_graph/*.smooth.final.gfa \
  -p /data/my_graph/giraffe_index

Graph Output

d) Aggregate Graphs from Individual Cohorts

WIP vg combine from vg toolkit looks promising.

2) Genomic background hashing for phenotype association of APOE locus

We extracted this locus around the APOE gene which will be the region we use for localized pangenome graph mapping. Pangenome graph mapping may then provide a genomic background to contexualize the high risk APOE alleles. This genomic background may capture trans expression effects and we will aim to code this using genomic hashing to represent different anonymized haploblocks that could be used in a federated manner across studies.

Focus on APOE Locus

downloaded from GWAS Catalog

wget https://ftp.ebi.ac.uk/pub/databases/gwas/summary_statistics/GCST013001-GCST014000/GCST013197/GCST013197.tsv.gz
gunzip GCST013197.tsv.gz

Extract Locus from Pre-constructed Pangenome Graph

WIP: odgi extract will be used, still need to figure out details.

Results

Conclusions

References