Statistical Genomics Foundry
Paper
domain/whole-genome-alignment

Li 2018 — Minimap2: pairwise alignment for nucleotide sequences

1. Citation

Li H. 2018. “Minimap2: pairwise alignment for nucleotide sequences.” Bioinformatics 34(18):3094–3100. doi:10.1093/bioinformatics/bty191. Preprint: arXiv:1708.01492 (v5, last revised 16 Mar 2018). Software man page consulted for the preset table: minimap2 manual at https://lh3.github.io/minimap2/minimap2.html, documenting version minimap2-2.28-dirty (r1237) (access date 2026-07-03).

2. Access note

The OUP published version is paywalled (all-rights-reserved). Content below is drawn from the arXiv preprint (1708.01492 v5) via the ar5iv HTML rendering plus the authoritative online man page (minimap2-2.28). The preprint body was read through ar5iv, which may drop equations/tables imperfectly — some formula terms marked accordingly. The preset→divergence numbers and per-preset -k/-w values come from the man page (v2.28), NOT the 2018 paper text; the paper text does not enumerate the asm5/asm10/asm20 presets (see §12). Mode: own-words paraphrase (restrictive posture); functional strings (preset names, flag values, %) kept verbatim as facts.

3. Thesis

Minimap2 is a single general-purpose alignment program that maps DNA or long mRNA against a large reference across regimes — accurate short reads, noisy long reads, spliced cDNA/Direct-RNA, and whole assembly contigs — while being several-fold to tens-of-fold faster than tools specialized in any one of those regimes at comparable or better accuracy.

4. Problem & context

New sequencing produces ultra-long reads (~100 kb average), full-length mRNA/cDNA reads, and assembly contigs over 100 Mb. Existing aligners were unable or too slow to process such data at scale. The paper positions minimap2 as the replacement for that gap and as a unifier across read types that previously each needed a specialized aligner.

5. Method / approach

  • Minimizer seeding. Seeds are minimizers — fixed-length k-mers chosen by a rolling hash as the smallest k-mer within each window. Reference minimizers are stored in a hash table keyed by the minimizer hash, value = list of reference locations. (Default seeding, no preset: -k 15, -w 10, i.e. 15 bp minimizers, window 10.)
  • Chaining. Colinear anchors (seed matches) are grouped into chains by dynamic programming, scoring matched bases and penalizing gaps with a concave gap cost. A chaining heuristic keeps up to h = 50 predecessor iterations.
  • Base-level (DP) alignment. Dynamic programming extends from chain ends and fills gaps between adjacent anchors, with a 2-piece affine (concave) gap cost that recovers long INDELs; default banding width ~500 bp.
  • Spurious-alignment heuristics. (a) Z-drop — break an alignment where the score drops too fast; (b) drop anchors that would imply a long insertion and long deletion (>~10 bp each) at the same position; (c) the 2-piece gap cost to recover longer INDELs.
  • Homopolymer-compressed (HPC) k-mers (optional, for noisy reads): collapse homopolymer runs before taking minimizers; used with -k 19 (map-pb uses -Hk19).
  • Divergence estimate. Approximates per-base error/divergence from seed statistics; ar5iv rendered the estimator as roughly ε̂ = (1/k)·log(n/m) [ar5iv-rendered form; verify against PDF]. The paper flags this estimate as “only approximate and can be biased,” with assumptions violated “especially when we take minimizers as seeds.”

6. Key claims / findings

  • Works across regimes: accurate short reads ≥100 bp; genomic reads ≥1 kb at error rate ~15%; full-length noisy Direct-RNA/cDNA; assembly contigs / near-complete chromosomes of hundreds of Mb.
  • Speed: 3–4× as fast as mainstream short-read mappers (Bowtie2, BWA-MEM) at comparable accuracy; ≥30× faster at higher accuracy for genomic and mRNA long reads.
  • Long-read (simulated human): >30× faster than BLASR/GraphMap/NGMLR; real ONT ~100 kb reads >70× faster than some aligners; ~200 CPU seconds on the simulated dataset.
  • Spliced (mouse cDNA): >40× faster than GMAP and SpAln; 94.0% of introns exactly correct.
  • Assembly (SMRT assembly vs GRCh38): ~20× faster than nucmer.
  • HPC k-mers raised SMRT overlap sensitivity from 90.9% to 97.4%.

7. Load-bearing statements (own-words paraphrase; functional strings verbatim)

  • On low identity, the paper states in effect that long exact seeds are rarely present anyway — paraphrase of “At low sequence identity, it is rare to see long seeds anyway.” (No numeric identity floor is given; see §9/§12.)
  • The divergence estimator is only approximate and can be biased, with its assumptions violated especially when minimizers are used as seeds (paraphrase).
  • Man page (v2.28), verbatim functional strings for the assembly presets:
    • asm5: “Typically, the alignment will not extend to regions with 5% or higher sequence divergence. Use this preset if the average divergence is not much higher than 0.1%.” Expansion: -k19 -w19 -U50,500 --rmq -r1k,100k -g10k -A1 -B19 -O39,81 -E3,1 -s200 -z200 -N50
    • asm10: “Use this if the average divergence is around 1%.” Expansion: -k19 -w19 -U50,500 --rmq -r1k,100k -g10k -A1 -B9 -O16,41 -E2,1 -s200 -z200 -N50
    • asm20: “Use this if the average divergence is around several percent.” Expansion: -k19 -w10 -U50,500 --rmq -r1k,100k -g10k -A1 -B4 -O6,26 -E2,1 -s200 -z200 -N50

8. Stated scope, assumptions, limitations

  • Targeted at DNA and long mRNA mapping against a large reference; long-read regime assumes error rate ~15% (genomic) / ~10% (map-ont default mode, per man page).
  • The divergence/error estimate is explicitly approximate and biased; not a calibrated identity measurement.
  • Preset guidance is expressed as average divergence recommendations, not hard validity bounds.

9. Failure modes / invalidity patterns

  • Low sequence identity → seed scarcity. As identity falls, long exact (or HPC) minimizer seeds become rare, so seeding sensitivity degrades. The paper names the mechanism qualitatively (“it is rare to see long seeds anyway”) but gives no numeric identity/divergence floor.
  • Preset/divergence mismatch (man page, v2.28). asm5 alignment “will not extend to regions with 5% or higher sequence divergence” — using asm5 on more-divergent input will truncate/miss divergent regions. Remedy is to move up the preset ladder (asm10 ~1%, asm20 ~several %) as average divergence rises.
  • Biased divergence estimate. The reported per-base divergence can be biased, particularly because minimizers (not all k-mers) are the seeds — do not treat it as an exact identity readout.
  • Z-drop / long-INDEL filters exist precisely because naive extension produces spurious alignments across INDELs; they are the named detectors for that failure.

10. What the source does NOT address

  • No stated numeric identity floor (e.g. the “~70% identity” figure some downstream skills cite) — neither the paper nor the man page gives such a threshold. Recorded as absent.
  • The 2018 paper text does not enumerate the named -x presets (asm5/asm10/asm20, map-ont, map-pb, map-hifi) — those live in the man page / software, not the paper body.
  • No treatment of protein alignment or short-read pair/insert-size modeling in depth here.

11. Open questions / ambiguities

  • The asm5/10/20 names: the man page only spells out the “5% or higher” extension ceiling for asm5; it does not verbatim state 10%/20% ceilings for asm10/asm20. The 10/20 in the names is a natural reading of an extension ceiling but is not stated verbatim — asm10/asm20 are described by average divergence (~1%, ~several %) instead. [summarizer-inferred that the numeric suffix tracks an extension ceiling; source does not confirm for asm10/asm20.]
  • Exact divergence estimator form uncertain from ar5iv rendering; confirm against the published PDF.

12. Guidance answers

  • What divergence does each asm preset target? From the man page (v2.28): asm5 — align does not extend past 5% divergence, recommended when average divergence is “not much higher than 0.1%”; asm10 — average divergence “around 1%”; asm20 — average divergence “around several percent.” These are man-page numbers, not paper-body numbers. Note the nuance: the named number (5/10/20) is an extension ceiling (explicit only for asm5), while the usage recommendation is a much lower average divergence — so “asm5 = 5% data” would misread the source; asm5 is recommended for ~0.1%-average data whose alignments won’t extend past 5%.
  • Identity/divergence floor below which seeding loses sensitivity? The paper states the mechanism (“At low sequence identity, it is rare to see long seeds anyway”) but gives no numeric floor. The often-cited ~70% identity cutoff is NOT present in the paper or the man page. Recorded as absent, not confirmed.
  • -k / -w defaults. No preset (map-ont default mode): -k 15, -w 10 (man page). Assembly presets (man page v2.28): asm5 -k19 -w19; asm10 -k19 -w19; asm20 -k19 -w10. map-pb expands to -Hk19 (HPC, k=19). Full preset expansion strings captured verbatim in §7.
  • Presets validated only up to a stated divergence ceiling? The only explicit ceiling statement is asm5’s “will not extend to regions with 5% or higher sequence divergence.” No general “validated up to X” claim is made for the others.
  • Version pinning. The preset table/divergence wording above is from minimap2-2.28-dirty (r1237). The 2018 paper (arXiv v5) predates and does not carry this table. A skill pinning “2.28+” is consistent with the man-page source used here.