For over two decades, biomedical research has relied on a single, linear sequence of DNA as the standard reference genome. While this coordinate system revolutionized genomics, it suffers from a fundamental flaw known as reference bias.
When a researcher sequences a new sample, algorithms align the resulting short or long sequencing reads to this linear reference. If the sample contains genetic variations—such as single-nucleotide polymorphisms (SNPs), insertions, deletions, or large structural rearrangements—that are absent from the reference, the alignment software struggles. Reads may be penalized for mismatches, mapped to the wrong location, or discarded entirely. As noted by Ting Wang et al. in their overview of The Human Pangenome Project: a global resource to map genomic diversity, this reliance on a single, linear sequence fails to capture the full spectrum of human genetic diversity, systematically masking complex variations across diverse populations.
To solve this, a paradigm shift has emerged: replacing linear reference strings with mathematical graphs that natively incorporate population-wide genetic variation.
The Anatomy of a Variation Graph
Instead of representing a genome as a single long string of nucleotides, a pangenome or variation graph represents a collection of genomes simultaneously.
Linear Reference: A T C G A T C G
\ /
Alternative Path: T T T
In this architecture, shared DNA sequences are stored as nodes, while genetic variations appear as diverging and re-joining paths (or "bubbles"). As Xiaoqin Huang et al. explain in Current status and prospects of super-pangenome studies in plants, this network structure allows for the simultaneous representation of SNPs, small indels, and large-scale structural variations across multiple individuals or even entire genera.
When a new genome is integrated, sequences matching the existing graph follow established paths, while novel sequences branch out as new nodes before tying back into the shared backbone. This approach allows researchers like Zhikun Wu et al. in Human pangenome analysis of sequences missing from the reference genome... to systematically capture millions of base pairs of "nonreference sequences" (NRSs) that play functional and evolutionary roles but are invisible on a standard linear scaffold.
However, moving from a 1D string to a complex multi-dimensional graph introduces significant cognitive and computational challenges. To make these structures interpretable, Wolfgang Beyer et al. developed Sequence tube maps: making graph genomes intuitive to commuters, a visualization framework inspired by public transit maps where individual genomes are depicted as continuous tracks running through shared sequence stations.
Why Graph Alignment Works: The Neutral Framework
The core reason graph-based references eliminate reference bias lies in the concept of a neutral mapping framework.
In a traditional alignment, a read originating from an alternative allele is penalized because it deviates from the reference. In a graph genome, the alternative allele is explicitly encoded as a primary path. As Cristian Groza et al. demonstrate in A unified framework to analyze transposable element insertion polymorphisms using graph genomes, representing structural variants as bifurcating paths or bubbles allows sequencing reads to align to multiple possible alleles without a built-in preference for a single "standard" sequence.
Traditional Linear Alignment:
Read: A T T T G C -> High penalty due to insertion gap
Reference: A — — — G C
Graph-Aware Alignment:
Read: A T T T G C -> Zero penalty; aligns perfectly to alternative path
Graph: A--[T T T]--G C
\_________/
By removing this penalty, graph-based alignment drastically improves performance in highly polymorphic and low-mappability regions—often called genomic "dark matter." Charles Markello et al. leveraged this in A complete pedigree-based graph workflow for rare candidate variant analysis, showing that graph-aware pipelines combined with deep learning models can accurately resolve complex loci to pinpoint rare disease-causing mutations that linear models miss.
This bias mitigation extends beyond basic variant calling into functional genomics. For example, in Graph Peak Caller: calling ChIP-seq peaks on graph-based reference genomes, Ivar Grytten et al. showed that mapping protein-DNA binding reads to a graph reference yields higher statistical confidence for epigenetic peaks located on non-reference variants, ensuring that functional genomic signals are not lost due to structural mismatches.
Resolving Zygosity and Structural Variations
Beyond simply finding where a read belongs, graph genomes change how we interpret complex genetic architecture.
Consider a large structural variant, such as a transposable element insertion. On a linear reference, determining whether an individual is homozygous or heterozygous for this insertion is notoriously difficult because reads spanning the insertion site fail to align cleanly. In a graph framework, the presence and absence of the insertion exist as parallel paths. As Cristian Groza et al. demonstrate, algorithms can determine zygosity by directly comparing the ratio of reads supporting the "presence" path versus the "absence" path.
This capability becomes vital when scaling up to population-level studies. By integrating high-quality, telomere-to-telomere (T2T) de novo assemblies into pangenome graphs, researchers can capture structural variants (SVs) that account for "missing heritability." Work by Heng Du et al. on pig pangenomes (Pangenome Reveals Gene Content Variations and Structural Variants...) and Glennis A. Logsdon et al. on human cohorts (Complex genetic variation in nearly complete human genomes) proves that embedding complete assemblies into a reference graph significantly enhances the accuracy of genotyping from standard, low-cost short-read sequencing data.
Scaling Up: Alignment-Free and Multi-Species Graphs
While graph alignment solves reference bias, mapping millions of reads through complex networks is computationally expensive. To address this, researchers have developed alignment-free graph genotyping.
Instead of performing full sequence alignment, tools like KAGE, developed by Ivar Grytten et al. (KAGE: fast alignment-free graph-based genotyping of SNPs and short indels), extract unique sequence fragments of length $k$ (k-mers) directly from the pangenome graph. Because the graph already contains the population's known variations, the index can identify population-unique k-mers ahead of time. Genotyping then becomes a fast, Bayesian counting exercise, accelerating variant calling by orders of magnitude without sacrificing accuracy.
On an evolutionary scale, graphs allow for reference-agnostic comparative genomics. Traditional methods align species back to a human or model scaffold, introducing severe cross-species bias. In Complete sequencing of ape genomes, DongAhn Yoo et al. utilized reference-agnostic pangenome graphs to perform all-to-all comparisons across multiple primate species, exposing lineage-specific innovations and structural divergences that linear mapping masks.
Similarly, in highly plastic bacterial genomes, the authors of Pangenome-spanning epistasis and coselection analysis via de Bruijn graphs (Juri Kuronen et al.) utilized colored de Bruijn graphs—where metadata tracks which specific strains contain which sequence nodes—to analyze epistatic interactions across both "core" and "accessory" genetic elements.
Abstracting the Graph: From Nucleotides to Genes
Graphs can also operate at different levels of abstraction depending on the biological question. While variation graphs track nucleotide-level changes, Heng Li et al. introduced "pangene" graphs in Exploring gene content with pangene graphs. Here, nodes represent entire genes rather than base pairs, streamlining the analysis of copy number variations, gene order alterations, and inversions across complex eukaryotic genomes.
This multi-tiered graph approach helps resolve a classic problem highlighted by Maximillian G. Marin et al. in Pitfalls of bacterial pan-genome analysis approaches...: differentiating between genuinely novel sequence insertions and the simple rearrangement of existing DNA segments. Meanwhile, in highly repetitive regions, frameworks like SEDEF by Ibrahim Numanagic et al. (Fast Characterization of Segmental Duplications in Genome Assemblies) demonstrate that separating small mutations from massive structural duplications within graph-based structures is essential for mapping complex evolutionary histories rapidly.
By shifting the foundational substrate of genomics from a single static string to a dynamic, multi-path graph, pangenomics ensures that genomic analysis is no longer biased toward an arbitrary reference, but is instead reflective of true biological diversity.
Key Foundational Work
Sequence tube maps: making graph genomes intuitive to commuters (2019) — contributed 1 concept elements to this direction.
Graph Peak Caller: Calling ChIP-seq peaks on graph-based reference genomes (2019) — contributed 1 concept elements to this direction.
Fast Characterization of Segmental Duplications in Genome Assemblies (2018) — contributed 1 concept elements to this direction.