Statistical Genomics Foundry
Paper
domain/pangenomicsdomain/whole-genome-alignment

Minigraph-Cactus pangenome graph construction (Hickey et al. 2024)

1. Citation

Hickey G, Monlong J, Ebler J, Novak AM, Eizenga JM, Gao Y, Human Pangenome Reference Consortium, Marschall T, Li H, Paten B. 2024. “Pangenome graph construction from genome alignments with Minigraph-Cactus.” Nature Biotechnology 42(4):663–673. DOI 10.1038/s41587-023-01793-w. Published in final edited form 2023 May 10. Access copy: PMC10638906 (NIHPA author manuscript, available in PMC 2024 Apr 17), https://pmc.ncbi.nlm.nih.gov/articles/PMC10638906/ . Accessed 2026-07-05.

2. Access note

Read the full PMC NIHPA author-manuscript text (main text + Methods) via the PMC HTML. Supplementary figures/tables (Supplementary Figs. 1–20, Supplementary Tables 1–5) are referenced but NOT included in this copy — any claim resting on them is marked below. The PMC copy is the accepted author manuscript, not the typeset Nature version.

License mode used: OWN-WORDS + functional strings verbatim (RESTRICTIVE)

The PMC record is an NIHPA author manuscript (“Nat Biotechnol. Author manuscript; available in PMC: 2024 Apr 17”), carrying “Reprints and permissions information is available at www.nature.com/reprints” and a PMC copyright notice — i.e. all-rights-reserved, NOT Creative Commons. Per the summarize-source contract, §7 and all load-bearing statements are rendered in own words. Exempt (reproduced verbatim as facts): tool/command/format names, numeric thresholds/defaults, versions, URLs, commit hashes, and short functional phrases.

3. Thesis

Minigraph-Cactus builds base-level pangenome graphs directly from whole-genome assembly alignments (not from variant calls), representing all variant scales in one model, and scales to 90 human haplotypes from the HPRC.

4. Problem & context

Reference coordinate systems introduce bias (unmappable reads in diverse samples; variant calls skewed toward the reference allele). Prior pangenome-graph construction faced a tradeoff: graphs built from reference-based variant calls (vg) could not represent nested variation and still carried some reference bias; graphs built by minigraph alone captured only SVs (≥50 bp) and were not usable for short-read mapping with then-current tools. Building directly from high-quality phased assemblies is equivalent to whole-genome multiple alignment — computationally very hard, so heuristics are required.

5. Method / approach

A five-step pipeline (steps runnable individually or as one workflow) that takes assemblies in FASTA and outputs a pangenome graph, a genome alignment, a VCF, and indexes for mapping with vg Giraffe. Steps:

  1. Minigraph SV graph construction — initial SV-only graph via minigraph; a “reference” assembly is chosen as an initial backbone and augmented with variation from the remaining assemblies in turn (POA-like iterative procedure). By default only variants affecting ≥50 bp are included. Minigraph does not collapse duplications (keeps gene copies separate → larger graph, fewer cycles).
  2. Minigraph contig mapping — each assembly (including the reference) mapped back to the SV graph with minigraph (generalized minimap2 seeding/chaining), producing base-level contig-to-graph alignments.
  3. Mapping filters / chromosome decomposition — filter spurious mappings; any mapping that would introduce a deletion edge of ≥10 Mb relative to the reference path is removed (tunable). Chromosome-assignment minimum thresholds: 75% for contigs ≤100 kb, 50% for 100 kb–1 Mb, 25% for >1 Mb; unmapped contigs labeled ‘unassigned’ and omitted.
  4. Base alignment with Cactus — minigraph mappings combined into a base-resolution graph using a modified Cactus base aligner (progressive multiple-alignment “pinch”/Cactus-graph machinery); the graph-collapsing step that scaled poorly with genome count was replaced with a POA approach that scales linearly (abPOA; max alignment length 10 kb).
  5. Indexing & clipping — combine chromosome graphs, reassign globally unique node IDs, collapse redundant nodes (gaffix), orient reference paths forward-only, then convert HAL→vg (hal2vg) and export VCF (vg deconstruct).

Three graphs may be produced during indexing:

  • Full graph — complete sequence, supports liftover; hard to index/map (unaligned centromeres); usually only an intermediate.
  • Default / clipped graph — clip out all stretches ≥10 kb that do not align to the minigraph SV graph. Rationale: large SVs absent from minigraph are treated as under-alignments, not true variants; the 10-kb threshold is arbitrary but empirically worked well and was sufficient to remove all centromeres (identified with dna-brnn) and matches abPOA’s max alignment length. Used for all results unless stated.
  • Allele frequency-filtered graph — remove all nodes present in fewer than N haplotypes (optional AF filter); improves short-read mapping and variant-calling accuracy; used for vg Giraffe mapping. Graph (3) ⊆ graph (2) ⊆ graph (1), node-ID compatible.

Outputs/formats (functional strings, verbatim):

  • Graphs output in GFA (version 1.1).
  • vg-native indexes: xg, snarls, and GBWT.
  • VCF via vg deconstruct (one site per snarl; uses the GBWT haplotype index to enumerate traversing haplotypes and compute phased genotypes).
  • Cactus natively emits HAL (hierarchical alignment); converted to vg via hal2vg. HAL feeds UCSC assembly hubs / cross-genome annotation but is not directly used for graph applications.

Tools/versions (verbatim): minigraph (base-level alignment since version 0.17); vg giraffe (version 1.37.0 in evaluation); BWA-MEM 0.7.17; hap.py 0.3.12 (docker jmcdani20/hap.py:v0.3.12); rtg vcf eval v3.91; Truvari version 3.5.0 (truvari collapse -r 500 -p 0.95 -P 0.95 -s 50 -S 100000); Picard 2.27.4 (LiftoverVcf, RECOVER_SWAPPED_REF_ALT); dna-brnn, gaffix, hal2vg. HPRC v1.0 graphs were built with an older version of the pipeline (satellite removed first via dna-brnn); exact HPRC-build commands pinned to Cactus commit 91bdd83728c8cdef8c34243f0a52b28d85711bcf. Code/binaries/Docker/manuals: https://github.com/ComparativeGenomicsToolkit/cactus .

6. Key claims / findings

  • Builds base-level pangenome graphs at the scale of “dozens to hundreds of vertebrate haplotypes”; demonstrated on 90 human haploid genomes from HPRC.
  • HPRC construction: from 47 HPRC diploid assemblies, held out HG002, HG005, NA19240; used the remaining 44 samples (88 haplotypes) plus two reference genomes GRCh38 and CHM13 (version 1.1) = 90 haploid genomes total. Ran the pipeline twice independently, once with GRCh38 as reference and once with CHM13.
  • The CHM13-based graph includes difficult/highly variant regions (e.g. acrocentric short arm of chr21) not represented in the GRCh38-based graph; slightly larger in sequence and in nodes/edges (Supplementary Table 1). In virtually all genome-wide mapping/variant-calling/ genotyping measures the CHM13-based graph was superior; with Giraffe/DeepVariant it improved on the GRCh38-based graph’s state-of-the-art accuracy.
  • Final pangenomes have roughly 200× more nodes and edges than the SV-only minigraph graph.
  • Drosophila melanogaster pangenome: 16 assemblies including reference dm6 (ISO1), 14 diverse strains + strain B7; sizes 132–144 Mb; demonstrates non-human applicability. Construction ~5 h for pangenomes (vs ~19 h for Progressive Cactus all-vs-all alignments); five D. mel graphs produced. DGRP short reads (205 inbred genomes, unrelated) used to evaluate mapping.
  • VCF export: bubbles/snarls decomposed to canonical SVs; each deconstructed-VCF allele specifies its path in the pangenome; variant frequency = number of assemblies traversing the path.

7. Load-bearing statements (OWN-WORDS paraphrase; functional strings verbatim)

  • The initial SV graph is reference-anchored: a “reference” assembly is chosen as an initial backbone and then augmented with variation from the other assemblies in turn (Methods, Minigraph SV graph construction).
  • Minigraph-Cactus requires at least one chromosome-level input assembly to be used as a reference backbone; graph quality/usefulness rises with the quality/completeness of all inputs (Discussion). [functional phrase “reference backbone” verbatim]
  • The reference coordinate system “can also be a source of bias” — bias appearing as unmappable reads in diverse samples, or, more subtly, variant calls being skewed toward the reference allele; pangenome graphs have been shown to reduce reference bias, but prior construction faced tradeoffs (Introduction/Discussion).
  • The choice of which included assembly serves as the reference backbone influences the topology and completeness of the graph; construction is dependent on the reference chosen (Discussion; HPRC results).
  • When an unaligned centromere is clipped, that region retains only the reference allele; other alleles are broken into two subpaths but are otherwise unaffected outside the clipped region (Fig. 1f caption).

8. Stated scope, assumptions, limitations (the source’s own caveats)

  • Requires at least one chromosome-level assembly as reference backbone (method assumption).
  • Results depend on the reference chosen; the two reference choices (GRCh38 vs CHM13) yield different graphs, and reference quality/completeness materially affects downstream performance.
  • The 10-kb clipping threshold is described as arbitrary (empirically effective, tied to abPOA max alignment length and centromere removal).
  • Minigraph includes only variants ≥50 bp by default in the SV graph; excluding small variation at that stage would prevent lossless path embedding — hence the Cactus base-alignment step.
  • Regions/sequence that cannot be confidently mapped (‘unassigned’, ‘minigraph-gap’ >100 kb, ‘clipped’ >10 kb) are removed; predominantly centromeric, acrocentric-short-arm, or segmental-duplication sequence enriched for putative assembly errors.
  • Closing note: “True graph-based genotyping formats and tools are needed” (Discussion) — VCF is a reference-relative interchange stopgap, not a native graph genotyping format.

9. Failure modes / invalidity patterns (referee-relevant)

  • Reference-anchoring is intrinsic: the graph is built on and clipped against a chosen reference backbone. In clipped (centromeric/unalignable) regions the graph keeps only the reference allele — non-reference sequence in those regions is not represented in the default graph. This is the paper’s own statement, not an external assertion of bias.
  • Reference-choice sensitivity: swapping the reference (GRCh38↔CHM13) changes graph topology, completeness, and which regions exist at all (e.g. chr21 acrocentric arm present only in CHM13). A single-reference build can therefore omit regions a better/other reference would include.
  • VCF export is reference-relative: vg deconstruct produces reference-anchored sites; the authors flag that VCF cannot represent nested variation and that true graph-native genotyping formats are still needed. Variant calls remain “skewed toward the reference allele” absent a graph-native representation.
  • Sequence dropped by filters: contigs below chromosome-assignment thresholds (75%/50%/25% by length) and stretches ≥10 kb unaligned to minigraph are omitted/clipped — sequence loss is a designed behavior, not an error; the amount removed is quantified only in supplementary figures (Suppl. Figs. 4, 17) not in this copy.
  • Large-deletion guard: mappings implying a deletion edge ≥10 Mb relative to the reference path are removed to avoid spurious nesting (all variants on a spanned arm being treated as nested in one large deletion) — a symptom the pipeline actively suppresses.

10. What the source does NOT address

  • No comparison to PGGB / seqwish / smoothxg. PGGB is not mentioned anywhere; the only head-to-head is against Progressive Cactus (star vs progressive alignment on Drosophila).
  • No cactus-pangenome command or CLI flag names. The paper describes the pipeline in prose (“five steps”); it does not give flag names such as --reference, --outDir, --outName, --vcf, --gfa, or --gbz. (Exact commands are deferred to the GitHub repo / doc/pangenome.md and the pinned Cactus commit.)
  • GBZ format is not mentioned. Indexes named are xg, snarls, and GBWT; GFA v1.1 is the graph output. (GBZ was introduced/became standard in the vg ecosystem around this time but this manuscript does not name it.)
  • No blanket claim that “non-reference haplotypes are systematically under-represented” — the paper frames Minigraph-Cactus as reducing reference bias while acknowledging residual reference-relative effects (clipping, VCF, reference choice).

11. Open questions / ambiguities

  • The exact CLI (cactus-* subcommands, argument names) is not in the paper — must be recovered from the repo at the pinned commit to author any runnable skill.
  • Whether the “current” pipeline (vs the older HPRC-v1.0 build) changes clipping/AF defaults is only partially addressed; the HPRC graphs used an older version with an extra dna-brnn satellite pre-removal step.
  • Quantitative sequence-loss figures live only in supplementary material not in this copy.

12. Guidance answers

  1. Reference-anchored or reference-free? — Reference-anchored. Own-words: the SV graph starts from a chosen “reference” assembly used as an initial backbone, augmented with the other assemblies; Discussion states the method requires at least one chromosome-level assembly as a reference backbone. (Functional phrase “reference backbone” verbatim.) “Reference-free” is never claimed.
  2. Variants relative to reference / reference-bias caveat? — Paths/variants are exported reference-relative (VCF via vg deconstruct; each allele specifies its pangenome path compared to the reference path). The paper’s OWN bias language: the reference coordinate system “can also be a source of bias,” manifest as variant calls being skewed toward the reference allele; it positions pangenome graphs as reducing reference bias while noting residual tradeoffs. The concrete under-representation the paper states: clipping unaligned/centromeric regions leaves only the reference allele there, and the choice of reference backbone influences graph topology and completeness (regions present under CHM13 absent under GRCh38). The paper does NOT assert a general “non-reference haplotypes are under-represented” caveat beyond these specific, reference-relative statements. (Rendered own-words; source is all-rights-reserved.)
  3. Output files + flags? — Formats confirmed: GFA (v1.1), VCF, HAL, plus vg indexes xg / snarls / GBWT. GBZ is NOT mentioned. Exact cactus-pangenome output-type flags (--vcf, --gfa, --gbz, --reference, --outDir, --outName) are NOT in the paper — it describes prose steps, deferring commands to the repo. GAP for flag names.
  4. Is --reference required? — The paper states a reference is required (at least one chromosome-level assembly as backbone), but the exact --reference flag name is not given.
  5. Scale demonstrated? — 90 human haploid genomes (44 HPRC samples = 88 haplotypes + GRCh38
    • CHM13); also 16 D. melanogaster assemblies.
  6. PGGB comparison / when to prefer each? — NOT addressed. No PGGB mention; only Progressive Cactus is compared (Drosophila). GAP.

Version pin recorded: minigraph ≥ 0.17; HPRC v1.0 graphs at Cactus commit 91bdd83728c8cdef8c34243f0a52b28d85711bcf; pipeline source at github.com/ComparativeGenomicsToolkit/cactus. No single “Minigraph-Cactus version” number is stated in the paper.