Panomap: Unbiased Nanopore Signal Mapping with Pangenome Variation Graphs

Technical Summary: Panomap – Unbiased Nanopore Signal Mapping with Pangenome Variation Graphs

Problem Statement
Nanopore sequencing generates raw ionic current signals that can be analyzed in "signal space" to enable real-time mapping and adaptive sampling decisions without the computational overhead of basecalling. However, existing signal-space mappers (e.g., RawHash2, Sigmoni) are designed around linear reference genomes. This reliance on a single linear reference introduces reference bias when the sample diverges from that reference. While pangenome variation graphs offer a compact, unbiased representation of genetic diversity by storing shared sequences once and encoding variants as alternative paths, current signal-space tools cannot utilize these graphs. Instead, they must treat pangenomes as collections of independent linear sequences (multi-FASTA), leading to redundant storage of shared sequences and separate, uncorrelated mapping results for reads matching multiple related references.

Methodology
Panomap is the first signal-space mapper designed to operate directly on pangenome variation graphs (in GFA format). The workflow consists of two primary phases: graph indexing and read mapping.

1. Graph Indexing:

   • Directional Expansion: Panomap converts the bidirected input graph into a directed graph by creating separate forward and reverse-complement copies of nodes, edges, and paths. This makes strand orientation explicit before indexing.
   • Per-Path Signal Seed Indexing: The tool walks each path in the directed graph. It "squigglizes" the base sequence into an expected event stream using a normalized pore-model table, applies event-level homopolymer compression (HPC), and tokenizes the events using a 4-bit adaptive quantizer. Seeds are extracted by hashing windows of consecutive tokens. Crucially, Panomap deduplicates seed records for shared graph nodes, ensuring that shared sequences are indexed only once, allowing the index to scale sublinearly with pangenome size.
   • Graph Coordinate Indexing: Two coordinate structures are built:
     • A path coordinate system that tags each node visit with a path identifier and cumulative base position, enabling colinear chaining within specific paths.
     • An approximate global 1D coordinate system (adapted from ODGI's 1D PG-SGD) used to detect and deduplicate overlapping chains across different paths.

2. Read Mapping:

   • Signal Processing: Raw nanopore signals are processed in chunks. The signal is converted to picoampere current, segmented into events, normalized, tokenized, and hashed into seeds matching the reference index space.
   • Anchor Formation and Chaining: Matches between read seeds and the seed index form graph anchors. Panomap expands these anchors into path-specific anchors using the path coordinate index. It then performs colinear chaining independently within each path using a formulation based on minimap2/RawHash2.
   • Chain Deduplication and Decision: Chains are scored and filtered. A global 1D coordinate system identifies and discards near-duplicate chains arising from different paths covering the same genomic region. A weighted decision score determines whether to emit a mapping based on the quality of the best chain relative to competitors, allowing for early decisions on short signal prefixes.

Key Contributions

• First Signal-Space Graph Mapper: Panomap bridges the gap between signal-space analysis and pangenome variation graphs, enabling direct mapping of raw signals to graph references.
• Efficient Indexing: By indexing paths within a graph and deduplicating seeds for shared nodes, Panomap achieves sublinear index scaling with respect to pangenome collection size, avoiding the redundancy inherent in multi-FASTA approaches.
• Path-Aware Chaining: The method projects graph anchors onto specific paths to perform chaining, preserving accurate nucleotide distances and handling cycles explicitly, which is critical for signal-space accuracy where downstream alignment correction is computationally difficult.

Results
The authors evaluated Panomap against state-of-the-art tools (RawHash2 and Sigmoni) across three scenarios:

1. Pangenome Scaling with a Good Reference: When a single reference already represents the sample well, Panomap maintains high mapping accuracy (F1 ~0.94 for SARS-CoV-2, ~0.89 for S. cerevisiae) as the reference collection grows from 1 to 8 sequences. In contrast, RawHash2 and Sigmoni showed regressed performance (lower F1 and recall) as the collection size increased, due to score dilution across multiple redundant targets.
2. Absent Strain Handling: When the exact sample strain was absent from the reference, Panomap improved mapping accuracy by incorporating related assemblies into the pangenome. Macro F1 scores increased as more related assemblies were added, whereas Sigmoni's performance declined.
3. Polymorphic Locus Recovery: At the highly polymorphic HLA-DRB1 locus, Panomap demonstrated superior recall for reads from divergent alleles (up to 41% divergence from the reference) when the source allele was missing from the reference but related alleles were present in the graph. This gain was most pronounced when mapping decisions were made from short signal prefixes (1–2 chunks), where linear references performed poorly.
4. Resource Efficiency: Panomap's index size scales sublinearly (e.g., growing only 2.2x for S. cerevisiae when the collection size increased 8x), whereas RawHash2's index grew near-linearly. Per-read processing time remained competitive, though Panomap was slightly slower on large, complex pangenomes (Zymo community) due to the cost of per-path chaining.

Significance
The paper claims that Panomap successfully brings "population-aware reference representation" into signal-space mapping. By operating directly on variation graphs, it eliminates reference bias without the storage redundancy of linear-reference collections. The results demonstrate that Panomap can leverage pangenome diversity to improve mapping accuracy for divergent samples and alleles, particularly in scenarios requiring rapid decisions from limited signal data. The authors identify graph-aware chaining and robust signal representation as key directions for future improvements in making signal-space mapping scalable and accurate for large-scale pangenomes.