Human Pangenome

The human pangenome is a graph-based reference representation of the human genome that integrates multiple phased, telomere-to-telomere (T2T) diploid assemblies from ancestrally diverse individuals into a unified sequence variation graph. For over two decades, clinical and population genomics relied on the linear human reference assembly (GRCh38), a mosaic haploid coordinate system derived predominantly (~70%) from a single donor (RP11) that leaves approximately 200 megabases (Mb) of heterochromatic, centromeric, and acrocentric sequence unresolved and omits thousands of polymorphic insertions, inversions, and variable-number tandem repeats (VNTRs) segregating across human populations. In 2023, the Human Pangenome Reference Consortium (HPRC) published its first draft human pangenome comprising 47 phased diploid assemblies (94 haplotypes) together with the T2T-CHM13 assembly, adding ~119 Mb of euchromatic polymorphic sequence and ~1,115 gene duplications absent from GRCh38.

By encoding shared ancestral segments as canonical sequence nodes and divergent haplotypes as alternative paths through a bidirected graph, the human pangenome eliminates reference mapping bias—the systematic failure of short-read and long-read aligners to map non-reference alleles when reads diverge from a single linear coordinate backbone. Graph-based pangenome representations enable quantitative topological characterization of complex structural variants (SVs), copy-number variations (CNVs), and hypervariable tandem repeat arrays at clinically critical loci including the major histocompatibility complex (MHC), lipoprotein(a) (LPA), cytochrome P450 CYP2D6, and the Intron 30 decamer VNTR of CACNA1C.

Limitations of the Mosaic GRCh38 Reference and T2T-CHM13

The International Human Genome Sequencing Consortium finished the initial euchromatic linear reference using hierarchical bacterial artificial chromosome (BAC) clones combined from multiple donors, with library RPCI-11 (RP11, an individual of African-European admixed ancestry) contributing ~70% of the sequence. Because a single linear string can represent only one haplotype at any genomic coordinate, GRCh38 introduces structural artefacts where BAC boundaries switch between distinct donor haplotypes, creating synthetic chimeric alleles and misassembled segmental duplications (SDs). Furthermore, GRCh38 contains ~151 Mb of gap placeholders (N-tracts) across centromeric alpha-satellite arrays (HORs), subtelomeric repeats, and the short arms (p-arms) of the five acrocentric chromosomes (Chr13, Chr14, Chr15, Chr21, and Chr22) housing nucleolar organizer regions (NORs) and 45S ribosomal RNA (rDNA) tandem arrays.

The Telomere-to-Telomere (T2T) Consortium resolved these linear gaps in 2022 by sequencing the effectively homozygous complete hydatidiform mole cell line CHM13 (46,XX) using ultra-long Oxford Nanopore Technologies (ONT >100 kb) reads and Pacific Biosciences (PacBio) high-fidelity (HiFi, ~15–20 kb at >99.9% accuracy) circular consensus reads scaffolded with Hi-C chromatin conformation capture. While T2T-CHM13v2.0 (including the HG002 Y chromosome) delivered a gapless 3.055-gigabase (Gb) assembly, any single haplotype still misses structural polymorphisms shared across global populations—typically ~20,000 to ~27,000 structural variants (≥50 bp) and ~4.2 million single-nucleotide variants (SNVs) per diploid human genome.

HPRC Phased Trio Assemblies and Graphical Fragment Assembly (GFA)

To capture global haplotype diversity, the HPRC utilized trio-binned PacBio HiFi, ultra-long ONT, and parental short-read k-mer phasing (Hifiasm and Verkko) to assemble maternal and paternal chromosomes independently without collapsing heterozygous structural variants. Across the initial 47 diverse individuals (94 haplotypes) and subsequent expanded cohorts (>350 diploid assemblies), the euchromatic fraction shows 99.99% base-level accuracy (quality value QV > 50) and resolves >99% of expected single-copy genes per haplotype.

These phased assemblies are unified into pangenome variation graphs in Graphical Fragment Assembly (GFA v1.1) format using three primary construction engines:

1. Minigraph (rGFA): Iteratively aligns assemblies against a reference backbone (GRCh38 or T2T-CHM13), recording structural variants ≥50 bp while preserving stable reference coordinates (SR tags).
2. Minigraph-Cactus: Extends Minigraph by adding base-level alignment (SNVs and small indels <50 bp) within structural variation bubbles while anchoring homologous chromosomes.
3. PanGenome Graph Builder (PGGB): An all-versus-all, reference-free construction pipeline combining wfmash whole-genome homologies, seqwish transitive closure of alignment matches, and smoothxg partial-order alignment smoothing, capable of modeling non-homologous acrocentric p-arm recombination (Chr13/14/15/21/22 Robertsonian translocation substrates).

Bidirected Sequence Graphs and Bubble Decomposition (Snarls)

Formally, a pangenome variation graph is defined as a bidirected sequence graph G = (V, E, P), where each vertex v ∈ V represents a DNA sequence segment of length |*v*| bp with a forward (+) and reverse-complement (−) orientation, each edge e ∈ E connects an oriented extremity of vertex u to an oriented extremity of vertex v, and P is the set of observed biological haplotype walks (W₁, W₂, ..., W_M) traversing the graph. Regions where haplotypes diverge from a conserved flanking anchor node s (source) and reconverge at a sink node t form nested topological structures termed ultrabubbles or snarls.

To quantify structural rearrangement density within a focal locus or snarl subgraph G_loc, graph-genomic pipelines compute the first Betti number (cyclomatic complexity, b₁):

b₁ = |*E*| − |*V*| + c

where |*E*| is the number of bidirected edges, |*V*| is the number of sequence nodes, and c is the number of connected components. In a strictly colinear locus containing only biallelic SNVs, each bubble adds two parallel edges and one alternative node (b₁ increases by 1 per biallelic site). In contrast, tandem repeat expansions, copy-number variations, and overlapping inversions generate recurrent backward loop edges (b₁ ≫ |*V*|), elevating cyclomatic density.

Haplotype Walk Shannon Entropy and Ollivier–Ricci Curvature

Because edge topology alone does not reflect how frequently different paths are traversed across human populations, allelic diversity across a snarl is quantified by the haplotype walk Shannon entropy (H_walk). Given M phased haplotypes partitioned across K distinct path topologies with empirical frequencies p_k = n_k / M (where ∑ p_k = 1), walk entropy is defined as:

H_walk = − ∑_*k*=1^K p_k log₂(p_k) (measured in bits)

Conserved coding exons and invariant CTCF anchor nodes exhibit H_walk → 0.0 bits, whereas hypervariable multi-allelic VNTR bubbles approach the theoretical ceiling log₂(M). For example, within the 30-bp decamer tandem repeat in Intron 30 of CACNA1C (Chr12p13.33), linear GRCh38 represents a single collapsed 300-bp array, whereas phased HPRC haplotypes span arrays from 3.2 kb to >30 kb with variable internal motif permutations, yielding H_walk = 6.17 bits across 94 haplotypes.

Furthermore, discrete geometric analysis of pangenome graphs utilizes Ollivier–Ricci curvature (κ_OR) across edges e = (u, v), comparing the Wasserstein-1 optimal transport distance (W₁) between neighborhood probability measures μ_u and μ_v against the shortest-path edge distance d(u, v):

κ_OR(u, v) = 1 − [W₁(μ_u, μ_v) / d(u, v)]

Conserved diploneme backbone nodes connected by dense local clique-like paths exhibit positive curvature (κ_OR > 0), whereas structural variant breakpoints, non-allelic homologous recombination (NAHR) hotspots, and high-copy VNTR boundary junctions act as geometric bottlenecks with strongly negative curvature (κ_OR < −0.50).

Eliminating Reference Bias and Genotyping Complex Structural Loci

When short sequencing reads (e.g., 150-bp Illumina reads) from a patient carrying a non-reference insertion or divergent haplotype are aligned to linear GRCh38 via BWA-MEM, reads spanning the variant breakpoint suffer alignment score penalties and either fail to map or misalign to paralogous loci (pseudogenes). Pangenome short-read aligners such as VG Giraffe index haplotype-consistent minimizer k-mers across the graph's embedded walks (GBWT—Graph Burrows–Wheeler Transform), mapping reads at speeds comparable to linear aligners while boosting recall of heterozygous insertions/deletions and enabling accurate copy-number genotyping (PanGenie) across medically essential pharmacogenes and immunogenetic regions:

• Cytochrome P450 CYP2D6 (Chr22q13.2): Resolves hybrid CYP2D6–CYP2D7 gene conversions and tandem duplications governing metabolism of >20% of clinical pharmaceuticals (tamoxifen, codeine, tricyclic antidepressants).
• Complement Component C4A/C4B (Chr6p21.33, MHC Class III): Disentangles long (HERV-K retroviral insertion) versus short cassettes and copy-number states (C4A copy number inversely correlates with systemic lupus erythematosus risk, and directly correlates with synaptic pruning in schizophrenia).
• Lipoprotein(a) LPA Kringle IV Type 2 (KIV-2): Resolves 5.5-kb tandem repeats ranging from 2 to >40 copies per allele that control plasma lipoprotein(a) concentrations and atherosclerotic cardiovascular risk.
• RHCE/RHD and SMN1/SMN2: Distinguishes paralogous sequence variants in spinal muscular atrophy (SMN1 exon 7 conversion) and complex Rh blood group antigens.

How much previously missing sequence does the human pangenome add to GRCh38?

Two complementary milestones expanded the human reference: first, the gapless T2T-CHM13 assembly resolved ~200 megabases (Mb) of centromeric alpha-satellites, segmental duplications, and acrocentric short arms (Chr13, 14, 15, 21, 22) that were represented as N gaps in GRCh38. Second, the HPRC first-draft pangenome (94 phased haplotypes across 47 diverse individuals) added ~119 Mb of euchromatic polymorphic sequence—roughly 90 Mb of which derives from structural variants, mobile element insertions, and variable-number tandem repeats (VNTRs) that segregate across global populations and cannot be represented in any single linear genome.

What is reference mapping bias, and how do variation graphs eliminate it?

Reference mapping bias occurs when sequencing reads derived from a non-reference allele accumulate mismatch or gap penalties when aligned against a single linear reference genome (GRCh38), causing them to map with lower quality scores or fail to align altogether. In heterozygous individuals, this systematically skews allele balance toward the reference allele during variant calling, chromatin immunoprecipitation (ChIP-seq), and allele-specific transcription (RNA-seq). Pangenome variation graphs embed common alternative alleles and structural variants directly as valid graph paths (GBWT haplotypes), allowing reads from non-reference haplotypes to align with zero penalty.

Why are acrocentric chromosomes represented as tangled subgraphs in reference-free pangenomes?

In humans, the short arms (p-arms) of the five acrocentric chromosomes (Chr13, Chr14, Chr15, Chr21, and Chr22) encode the 45S ribosomal RNA (rDNA) arrays and share massive megabase-scale blocks of >99% identical satellite and segmental duplication sequences (pseudo-homologous regions). When assembled in a reference-free pangenome graph (PGGB), these five p-arms collapse into a connected, high-cyclomatic-complexity community subgraph, reflecting ongoing heterologous recombination and Robertsonian translocation exchange among all five acrocentric chromosomes during human meiosis.

1. Liao, W.-W., Asri, M., Ebler, J., et al. (2023). "A draft human pangenome reference." Nature, 617(7960), 312–324. https://doi.org/10.1038/s41586-023-05896-x
2. Nurk, S., Koren, S., Rhie, A., et al. (2022). "The complete sequence of a human genome." Science, 376(6588), 44–53. https://doi.org/10.1126/science.abj6987
3. Sirén, J., Monlong, J., Chang, X., et al. (2021). "Pangenomics enables genotyping of known structural variants in 5202 diverse genomes." Science, 374(6574), abg8871. https://doi.org/10.1126/science.abg8871
4. Paten, B., Novak, A. M., Eizenga, J. M., & Garrison, E. (2017). "Genome graphs and the evolution of genome inference." Genome Research, 27(5), 665–676. https://doi.org/10.1101/gr.214155.116