Pangenome

Unlike a linear reference genome, a pangenome is represented as a sequence graph or variation graph. Nodes represent sequence segments, and edges define allowed paths through the genome. This structure explicitly encodes structural variants (SVs) —insertions, deletions, inversions, and duplications larger than 50 base pairs—that are invisible to linear references.

• Directed Acyclic Graph (DAG): A common topology where edges flow in one direction, preventing cycles.
• Bidirected Graphs: Used to represent both forward and reverse-complement strands simultaneously.
• Node Labels: Store the actual nucleotide sequence for each segment.

This graph-based model eliminates reference bias, where reads from structurally divergent regions fail to align to a single linear reference.

Pangenomes partition genes into three categories based on their frequency across a species:

• Core Genome: Genes present in all individuals of the species. Typically encodes essential housekeeping functions like ribosomal proteins and central metabolism.
• Accessory Genome: Genes present in a subset of individuals (e.g., 15-95% frequency). Often includes virulence factors, antibiotic resistance genes, and niche-adaptive traits.
• Cloud Genome: Rare genes found in very few individuals (<15%). Represents the species' genetic reservoir for rapid adaptation.

This partitioning is critical for understanding horizontal gene transfer in bacteria and presence-absence variation (PAV) in plants and eukaryotes.

Aligning sequencing reads to a pangenome requires specialized algorithms that navigate the graph topology. Unlike linear alignment, the mapper must consider all possible paths through the graph simultaneously.

• Seed-and-Extend: Identifies exact matches (seeds) between the read and graph nodes, then extends alignments across edges.
• POA (Partial Order Alignment): A dynamic programming method that aligns a sequence to a directed acyclic graph.
• GAF (Graph Alignment Format): A standardized output format that records the path of nodes and orientations traversed by each read.

Tools like vg (variation graph toolkit) and GraphAligner perform this mapping, producing alignments that capture allele-specific expression and haplotype-resolved variants.

Storing the full genomic diversity of a species as an uncompressed graph is computationally prohibitive. Several compression strategies reduce the memory footprint:

• GBWT (Graph Burrows-Wheeler Transform): A compressed index of all haplotypes in the graph, enabling efficient read alignment and haplotype querying.
• r-index: A run-length compressed suffix array that scales sublinearly with the number of haplotypes.
• Minimizer Indexing: Reduces the search space by indexing only a sparse subset of k-mers (minimizers) from the graph.
• GFA (Graphical Fragment Assembly) Format: A lightweight text format for representing sequence graphs, supporting both segment and link lines.

These techniques enable population-scale pangenomes like the Human Pangenome Reference Consortium (HPRC) to be queried on commodity hardware.

Machine learning models require fixed-dimensional numerical inputs, but pangenome graphs have variable topology. Graph embedding techniques convert nodes, edges, and paths into dense vector representations:

• Node2Vec: Learns embeddings by performing biased random walks on the variation graph, capturing both structural equivalence and homophily.
• Graph Neural Networks (GNNs): Message-passing architectures that aggregate features from neighboring nodes, learning representations that encode local graph topology.
• Poincaré Embeddings: Hyperbolic space embeddings that naturally capture the hierarchical, tree-like structure of gene families within the pangenome.
• PanGenome Graph Neural Network (PanGNN): A specialized architecture that learns allele-specific embeddings by propagating information along haplotype paths.

These embeddings enable downstream tasks like variant effect prediction and gene presence-absence classification directly from graph structure.

Standard k-mer tokenization fails to capture structural variants because it operates on a single linear sequence. Pangenome-aware tokenization strategies include:

• Path-Based Tokenization: Tokenizes each haplotype path through the graph as a distinct sequence, generating embeddings that are aware of bubble structures (regions of variation).
• Anchor K-mers: Identifies k-mers that are unique and conserved across all haplotypes, using them as stable coordinate anchors while allowing variable-length gaps between them.
• Graph K-mers: Extends the k-mer concept to graphs by enumerating all possible k-length walks starting from each node.
• Bidirectional Tokenization: Encodes both the forward and reverse-complement orientations of graph nodes, respecting the double-stranded nature of DNA.

These methods ensure that presence-absence variation and copy number variation are explicitly represented in the tokenized input to downstream models.

The method of embedding named haplotypes as distinct paths threaded through the shared graph backbone. Each individual's diploid genome is stored as a pair of walk coordinates rather than a separate sequence.

• GBWT (Graph BWT): A compressed index storing thousands of haplotypes as paths through the graph nodes.
• Enables rapid population-scale genotyping without realignment.
• Allows the graph to serve as both the reference and the variation catalog simultaneously.
• Critical for storing the diversity of large cohorts like the Human Pangenome Reference Consortium (HPRC).

A bubble is a directed acyclic subgraph representing a site of genetic variation where multiple alternative alleles diverge and rejoin the main path.

• SNV Bubble: A single nucleotide polymorphism creates a simple two-node branch.
• SV Bubble: Large insertions or deletions create long, complex alternative paths.
• Nested Bubbles: Variation within variation, common in highly diverse regions like the Major Histocompatibility Complex (MHC).
• Embedding algorithms must learn to map reads to the correct bubble arm based on context.

A stable, unambiguous addressing scheme for positions within a pangenome graph. Linear coordinates (chr:pos) fail because insertions shift coordinates between haplotypes.

• Node ID + Offset: The most basic system, but fragile to graph edits.
• Stable Sequence IDs: Hashing sequence content to generate position-independent identifiers.
• PanSN Naming: A convention prefixing coordinates with the haplotype or assembly name (e.g., CHM13#chr8:1000).
• Essential for annotating genes and regulatory elements consistently across diverse assemblies.

Applying GNNs directly to the pangenome graph structure to learn node and edge embeddings that capture evolutionary and functional constraints.

• Node Features: Initialized with sequence k-mer content or conservation scores.
• Message Passing: Aggregates information from neighboring nodes, allowing the model to see flanking context across bubble boundaries.
• GraphSAGE or GAT architectures can predict variant pathogenicity or regulatory impact.
• Bypasses the need to linearize the graph, preserving the topological relationships of structural variants.