Building, understanding, and using pangenomes

Erik Garrison

​

University of Tennessee (UTHSC), Memphis

​

@Workshop on Genomics, Český Krumlov

January 13, 2025

Key model: variation graph = nodes, edges, paths

Key model: variation graph = nodes, edges, paths

Embeds genomes in a common sequence graph

New ideas often have a long history

St. Agatha pipetting a biosample into a nanopore sequencer

c. 1420

This all seems cool and “new” but ideas are rarely that.

​

Pangenomes and variation graphs have a long^† history.

​

(^†for genomics)

pipette

biosample

nanopore sequencer

nine versions Valerio Magrelli’s

poem “Campagna Romana” (1981)

variation graphs: not a new idea!

https://doi.org/10.1007/978-3-030-38281-0

Group B Streptococcus assemblies from 2002

Pangenome: not a new concept

First collections of multiple genomes from the same species demonstrated substantial differences.

​

This was unexpected and required new theory to understand.

​

A single reference is not enough to explain genomic diversity. Even many genomes may not be enough.

​

Some genes are shared among all individuals: these are “core”, while others are not—we call them “accessory”.

Lessons from language modeling: Heaps’ law

​

A pangenome is:

​

Closed: our observations of new genes with new genomes diminish.

​

Open: we continue to see new genes as we add more genomes.

https://doi.org/10.1007/978-3-030-38281-0

The exponent α determines whether the pangenome is open (α ≤ 1) or closed (α > 1). The top panel shows data for an open pangenome species, P. marinus; the bottom panel for a closed pangenome species, S. aureus

Pangenome research timeline

2000-2010s: counting genes

​

~2015: let’s take it to the sequence level (genome graphs)

​

2020s: complete assemblies (T2T pangenomes)

github.com/vgteam/vg

“Computational pan-genomics: status, promises and challenges” https://doi.org/10.1093%2Fbib%2Fbbw089

yup… we can generalize most standard bioinformatic algorithms to graphs, as in Partial Order Alignment →

Wait! You can align sequences to graphs?

https://doi.org/10.1093/bioinformatics/18.3.452

reads aligned to a variation graph

vg resolves reference bias at known indels in HG002

50x 2x150bp Illumina sequencing of HG002

Using variation graphs to observe CCR5-delta in ancient samples

Rui Martiniano

Exonic deletion in the HGSVC dataset correctly genotyped by vg

reads vs. graph

reads vs. GRCh38

vg giraffe: approach

vg giraffe

is accurate

enough

vg giraffe is very fast

vg giraffe

improves

variant calling

vg giraffe lets us scale: PCA from SVs in 5k genomes

Building a draft

human pangenome

Q1 2025

Sample selection was constrained by:

• trio status in Coriell biobank (-Europeans)
• low cell line passage count (--Europeans)
• genetic diversity (+++Africans)
• drift (+Asians, ++Americas)

Draft pangenome composition

Amazing assemblies approach reference quality

Haplotype-resolved assemblies from trio-hifiasm.

​

They are really good, according to realignment of reads to the assemblies and model of assembly completeness—nearly as good as T2T-CHM13!

Mobin Asri

Then we made 5 pangenome (reference) graphs…

Minigraph

https://doi.org/10.1186/s13059-020-02168-z

Minigraph-Cactus

https://doi.org/10.1038/s41587-023-01793-w

https://doi.org/10.1101/2023.04.05.535718

Key conceptual differences between HPRC pangenome construction methods

​

minigraph: just SVs, no complex stuff, one reference.

​

minigraph-cactus: add SNPs, clean up the breakpoints, useful for alignment, one reference.

​

pggb: everything-vs-everything, hard to align to, useful for studying evolution and pangenome structure at all scales, all genomes are references.

​

“Collapse” in high-copy repeats →

minigraph-Cactus creates a hierarchical pangenome rooted in the reference genome, ensuring compatibility with standard tools.

PGGB creates graphs in which each genome can act as a reference, so we choose our reference as needed by later analysis or work in graph space.

Why do we need a human pangenome?

T2T-CHM13 adds ~200MB of heterochromatin to reference.

​

Draft pangenome adds ~100MB of polymorphic euchromatin (and a lot more heterochromatin),

​

0.6-4.4 Mb of additional genic sequences per haplotype compared to GRCh38 (38 gene CNVs/haplotype).

Pangenome growth (PGGB)

PGGB: all chromosomes, layout with path-guided SGD

Simon Heumos

C4A/B in pggb graph

MHC class II

6p

MHC

centromere

C4A/B in pggb graph

Christian Fischer (UTHSC)

MHC class II

https://github.com/chfi/gfaestus

Genbank annotations on top of the pggb HPRC graph.

C4A/B in pggb graph

Christian Fischer (UTHSC)

C4A/B

MHC class II

https://github.com/chfi/gfaestus

Genbank annotations on top of the pggb HPRC graph.

C4A/B in pggb graph

Christian Fischer (UTHSC)

copy number is related to schizophrenia risk

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4752392/

We learn that genome evolution is often nonlinear

complement component 4 locus

Large SVs predominantly occur at VNTRs which are simply loops in our pggb graphs.

We learn that genome evolution is often nonlinear

C4A

C4A

C4B

C4B

C4B

Amylase: how humans evolved a taste for agriculture

Amylase breaks starch into simpler sugars

48

Amylase is a multi-copy gene family

Human amylase copy number diversity

50

28 amylase structural haplotypes

11 consensus structures

Deconvolving haplotypes from short reads

51

Recent evolution from 533 ancient European genomes

Data from Allentoft et al. 2024, Nature

Evidence for selection of high-copy amylase haplotypes

Allele frequency changes over last ~12,000 years in eurasian populations.

Evidence for selection of high-copy amylase haplotypes

Allele frequency changes over last ~12,000 years in eurasian populations.

Bolognini*, Halgren*, Lou*, Raveane*, Rocha* et al., Nature

Recombination between heterologous acrocentric chromosomes

Workflow

https://github.com/human-pangenomics/HPP_Year1_Assemblies

HPRC assemblies

Mapping against the whole CHM13

PanGenome Graph Builder (PGGB)

https://github.com/pangenome/pggb

Acrocentric contigs covering (+/- 1Mbp) both the p and q arms (pq-contigs)

+ HG002 contigs >= 300kbps� which map to acrocentrics

Untangling the pangenome graph

We look into the graph from the perspective of chromosome 13. Full information from pangenome plus reference annotations.

Untangling the pangenome graph

We look into the graph from the perspective of chromosome 13. Full information from pangenome plus reference annotations.

Untangling the pangenome graph

We look into the graph from the perspective of chromosome 13. Full information from pangenome plus reference annotations.

Untangling the pangenome graph

We look into the graph from the perspective of chromosome 13. Full information from pangenome plus reference annotations.

64

Pseudo Homologous Regions

14

21

FISH probes from GRC clones

Guarracino et al., 2023, Nature

Complete sequence of Robertsonian chromosomes

de Lima*, Guarracino* et al., in preparation

GM03786

GM04890

t(13;14)

t(14;21)

GM03417

Pseudo-homologous regions (PHRs)

Guarracino et al., 2023, Nature

https://doi.org/10.1007/BF02910454

https://doi.org/10.1016/j.ajog.2004.02.062

Why was it not possible to see this before using traditional laboratory techniques?

​

← 300 FISHing experiments to find one heterologous synapse.

Implicit pangenomes

Heumos et. al. https://doi.org/10.1093/bioinformatics/btae363

Pangenome graph of the HPRCy1 including 88 haplotypes.

​

Represents all variation of all types between all genomes.

Pangenome graphs = Pangenome alignments

Alignments can be represented as sets of matches, which are just pairs of positions in two genomes.

Garrison & Guarracino 2022 https://doi.org/10.1093/bioinformatics/btac743

Pangenome graphs = Pangenome alignments

To build the graph, we condense matched genome letters into single nodes.

Garrison & Guarracino 2022 https://doi.org/10.1093/bioinformatics/btac743

Pangenome graphs = Pangenome alignments

Variation graph represents all alignments and all genomes.

Garrison & Guarracino 2022 https://doi.org/10.1093/bioinformatics/btac743

Is this the end of genomics?

One model:

• all parts of all genomes
• all types of variation
• genomes of different species, why not!
• put it all in there!

​

GENOME

ALL

Y’ALL

Is this the end of genomics?

One model:

• all parts of all genomes
• all types of variation
• genomes of different species, why not!
• put it all in there!

But… there are problems…

GENOME

ALL

Y’ALL

1.05 Terabytes

One chromosome.

You’re not alone agolicz.

You’re not alone Wwwwwwwyc.

😭😭😭

Don’t cry, wjwei-handsome

Building pangenome graphs is hard…

Why are graphs hard?

Graphs are “sticky”...

Anything can connect to anything else.

Nonlinear relationships = you never know what you need to know.

That means you need to remember everything, all at once.

​

==

S. cerevisiae pangenome graph

graph adjacency matrix

(node:node edges)

from Yue et. al. 2017 https://doi.org/10.1038/ng.3847

​

Why are graphs hard?

Graphs are “sticky”...

Anything can connect to anything else.

Nonlinear relationships = you never know what you need to know.

That means you need to remember everything, all at once.

​

Why are graphs hard?

Graphs are “sticky”...

Anything can connect to anything else.

Nonlinear relationships = you never know what you need to know.

That means you need to remember everything, all at once.

We cheated when we made the HPRCy1 graphs.

Yup.

We broke it up by chromosome.

🙈

🙊

Why does it matter if we cheated?

​

Guarracino et al. 2023 is why.

​

Heterologous acrocentric chromosomes recombine.

Recombination occurs here

Let’s overcomplicate things

→

→

alignments

pangenome graph

… “untangling”

from Guarracino et al. 2023.

Let’s overcomplicate things

→

→

alignments

pangenome graph

untangling means

… alignments?

The graph vanishes

→

→

alignments

pangenome graph

… alignments?

X

Implicit data structures

Useful pangenome queries:

• liftover (coordinate projection)
• subgraph extraction (can be used to divide-and-conquer graph build)
• local haplotype matching, homology (untangling)
• variant calling
• conservation
• incomplete lineage sorting

But, we want to avoid building the whole graph up front to get these.

We can obtain these with alignments and an implicit representation of the graph.

Implicit interval tree

https://doi.org/10.1093/bioinformatics/btaa827

100 150

130 200

170 300

180 250

200 250

250 350

270 300

300 320

350 450

390 420

==

IIT

IMPG: IMplicit Pangenome Graphs

Alignment [start, end) → intervals → implicit interval tree (per genome)

wfmash

https://github.com/pangenome/impg

https://github.com/waveygang/wfmash

Querying impg

1. Collect ranges overlapping a target range.

Querying impg

• Collect ranges overlapping a target range.

Querying impg

• Collect ranges overlapping a target range.

Querying impg

• Collect ranges overlapping a target range.

Querying impg

• Collect ranges overlapping a target range.
• For each collected range, walk the cigar to translate coordinates, and goto (1) …

Querying impg

• Collect ranges overlapping a target range.
• For each collected range, walk the cigar to translate coordinates, and goto (1) …

Querying impg

• Collect ranges overlapping a target range.
• For each collected range, walk the cigar to translate coordinates, and goto (1) …

Querying impg

• Collect ranges overlapping a target range.
• For each collected range, walk the cigar to translate coordinates, and goto (1) …

Querying impg

• Collect ranges overlapping a target range.
• For each collected range, goto (1) …
• Deduplicate

​

Result is the “transitive closure” of aligned ranges in pangenome. If graph == alignments, a subgraph query.

Telomere-to-telomere great ape assemblies

Homo sapiens (human)

Symphalangus syndactylus (siamang gibbon)

Gorilla gorilla (gorilla)

Pongo pygmaeus (Bornean orangutan)

Pongo abelii (Sumatran orangutan)

Pan troglodytes (chimpanzee)

Pan paniscus (bonobo)

phylogeny from PGGB graph on hsa_chr6

Telomere-to-telomere great ape assemblies

Homo sapiens (human)

Symphalangus syndactylus (siamang gibbon)

Gorilla gorilla (gorilla)

Pongo pygmaeus (Bornean orangutan)

Pongo abelii (Sumatran orangutan)

Pan troglodytes (chimpanzee)

Pan paniscus (bonobo)

phylogeny from PGGB graph on hsa_chr6

Telomere-to-telomere great ape assemblies

Homo sapiens (human)

Symphalangus syndactylus (siamang gibbon)

Gorilla gorilla (gorilla)

Pongo pygmaeus (Bornean orangutan)

Pongo abelii (Sumatran orangutan)

Pan troglodytes (chimpanzee)

Pan paniscus (bonobo)

phylogeny from PGGB graph on hsa_chr6

1 graph.

35 days.

job failed

Telomere-to-telomere great ape assemblies

Homo sapiens (human)

Symphalangus syndactylus (siamang gibbon)

Gorilla gorilla (gorilla)

Pongo pygmaeus (Bornean orangutan)

Pongo abelii (Sumatran orangutan)

Pan troglodytes (chimpanzee)

Pan paniscus (bonobo)

phylogeny from PGGB graph on hsa_chr6

1 graph.

35 days.

job failed

New implicit graph approach.

Align all-vs-all

→ impg index (subgraphs)

→ alignment-based analysis (MAF)

MHC queried from impg

Polymorphic inversion on human 8p23.1

inversion

5’ flank

3’ flank

defensin tangle

Conservation scores show “hotspots” of fast evolution

Conservation scores broadly reflect constraint

ILS: Chr 6 (chm13 ref frame)

Dots = mean quartet voting for 100 kb window

Lines = sliding mean 3 Mb

Ref Length:

172,126,629

​

QI-sites: 616,403

(around purple)

ILS: Chr 6 (mPonAbe1 ref frame)

Dots = mean quartet voting for 100 kb window

Lines = sliding mean 3 Mb

Ref Length:

172,605,364

​

QI-sites: 634,168

(17,765 more)

​

Andrea Guarracino (pggb, wfmash, seqwish, odgi, chromosome communities)�Simon Heumos (pggb, odgi)�Flavia Villani (pggb, applications to mouse, popgen)�Njagi Mwaniki (wfmash, WFA applications)�Santiago Marco-Sola (WFA, wfmash)�Pjotr Prins (guidance, vcflib, vcfwave)�Richard Durbin (guidance)�Nicole Soranzo (guidance, support)�Benedict Paten (vgteam)�Hao Chen (rat, mouse)�Zhigui Bao (applications)�Lorenzo Tattini (yeast pangenomes)�Enza Colonna (applications to mouse, popgen)�Nadia Pisanti (algorithms)�Luca Pinello (applications)�Jennifer Gerton (robertsonians)�Adam Phillippy (robertsonians)�Peter Sudmant (primate pangenomes)�Robert Williams (guidance)

HPRC pangenomes working group and many others

funders include

Tennessee, NSF, NIH

Practical!

Let’s build some pangenome variation graphs with pggb!

First: a deeper dive into how the method works.

Then: we’ll work through small examples to learn how to drive it.

wfmash (biWFA)

PanGenome Graph Builder

seqwish (unbiased graph builder)

PanGenome Graph Builder

smoothxg (graph normalization)

PanGenome Graph Builder

wfmash makes initial alignments

the wavefront algorithm (WFA)

Needleman-Wunsch

high-order WFA (WF-lamba)

wfmash makes initial alignments

the wavefront algorithm (WFA)

Needleman-Wunsch

high-order bidirectional WFA (BiWFλ)

seqwish builds the graph

Path-guided stochastic gradient descent algorithm to optimize 1D order to best-match positions in embedded paths.

Pangenome graph with 12 ALT sequences of the HLA-DRB1 gene from the GRCh38 reference genome.

smoothxg organizes & normalizes the graph

… then we run MSA over this, locally

odgi helps us understand the pangenome

https://www.biorxiv.org/content/10.1101/2021.11.10.467921v1

identity

position

orientation

copy number variation

2d layout

ODGI is meant to be a basic toolkit for interacting with pangenome graphs.

​

It uses the embedded genomes as references.

https://doi.org/10.1101/2023.04.05.535718

Putting it all together!

Test material today

1. A few genes from HLA-D (MHC class II) in humans — getting started
1. https://github.com/pangenome/pggb-workshop/tree/evomics2025
2. Yeast chromosome 6 — scaling up
• ~/workshop_materials/pangenomics/cerevisiae.chrV.fa
• you will want to apply samtools faidx to this… pggb will warn you
3. Whole yeast chromosomes — looking at chromosome variation
• ~/workshop_materials/pangenomics/cerevisiae.pan.fa.gz

Example: yeast chromosome 6

Yue, JX., Li, J., Aigrain, L. et al. Contrasting evolutionary genome dynamics between domesticated and wild yeasts. Nat Genet 49, 913–924 (2017). https://doi.org/10.1038/ng.3847

pafplot of

initial alignments from wfmash

a bit of the 2D layout

you have to zoom in in the web browser

path view

position view

inversion view

copy number view

TBC1D3

for Evan Eichler

chr17 in PGGB HPRCv1

sequence clipped by minigraph-cactus

annotation of clipped sequences in minigraph-cactus for A. thaliana pangenome