Source-linked AI summary
New strategies to improve minimap2 alignment accuracy
Heng Li
TL;DR
Earlier minimap2 versions could misalign reads in highly repetitive regions and fail to align long INDELs reliably. Minimap2 v2.22 introduces additional anchoring, rechaining, and structural-variant-aware rescoring, with improved mapping and long-INDEL sensitivity while maintaining similar performance to older minimap2.
Problem
Earlier minimap2 versions could misalign reads in highly repetitive regions and break or fail alignments across long INDELs.
Method
Minimap2 v2.22 adds high-occurrence minimizer anchors, minigraph-based rechaining, and alignment rescoring that penalizes long structural variants more mildly.
Results
Minimap2 v2.22 matched Winnowmap2 on simulated SV mapping, was more sensitive at comparable specificity for real INDELs longer than 1kb, and retained similar performance to older minimap2.
Takeaways & Limitations
The improvements make minimap2 more reliable for repetitive-region mapping, structural variants, and contiguous alignment across long INDELs without substantially changing runtime.
Abstract
from arXiv · showhide
Summary: We present several recent improvements to minimap2, a versatile pairwise aligner for nucleotide sequences. Now minimap2 v2.22 can more accurately map long reads to highly repetitive regions and align through insertions or deletions up to 100kb by default, addressing major weakness in minimap2 v2.18 or earlier. Availability and implementation: https://github.com/lh3/minimap2
1 INTRODUCTION
Minimap2 v2.18 and earlier had difficulty mapping reads in highly repetitive regions, especially with structural variation, and aligning long INDELs or through VNTRs.
- Earlier minimap2 versions sometimes misaligned reads from highly repetitive regions because they ignored high-occurrence seeds.Reads with structural variations could also be misplaced in these regions.
- Older minimap2 could not efficiently align long insertions or deletions and often broke alignments around VNTRs.
2 METHODS
Minimap2 v2.22 adds anchors in sparse regions, improves chaining across long INDELs, and rescoring for structural variants while retaining practical runtime.
- 2.1 Rescuing high-occurrence k-mers: Reads with few low-occurrence minimizers could fail chaining because the original mapper selected only low-occurrence minimizers during mapping.Minimap2 retains all minimizers during indexing, but applies a few-hundred occurrence cutoff for long-read mapping against a human genome.
- 2.1 Rescuing high-occurrence k-mers: When adjacent low-occurrence k-mers are at least 500bp apart, v2.22 adds lowest-occurrence minimizers between them as additional anchors.It selects floor(|x1 − x2|/500) minimizers with a binary heap, increasing alignment time by a few percent on real data.
- 2.2 Aligning through longer INDELs: The original 500bp chaining band could stop at longer gaps, while its long-join heuristic could fail around VNTRs because short chains often overlap.
- 2.2 Aligning through longer INDELs: Early escape after 50 unimproved iterations could break chaining around copy-number changes in long segmental duplications, even with a large band.
- 2.2 Aligning through longer INDELs: V2.22 uses DP chaining for short chains and a ported minigraph rechaining algorithm for longer INDELs, replacing the less reliable long-join heuristic.The rechaining step can resolve overlaps between short chains, although the minigraph algorithm is slower for long-read mapping.
- 2.3 Properly mapping long reads with SVs: V2.22 rescored alignments using matching bases, substitutions, and gap opens so long structural variants receive a milder penalty than under affine-gap scoring.The scoring approximates per-base sequence divergence with a minimum value of 2%, and rescoring is linear-time with negligible practical cost.
3 RESULTS
Minimap2 v2.22 retained high mapping accuracy, improved consistency in repetitive regions and long-INDEL alignment, and preserved runtime close to older minimap2 versions.
- Both minimap2 versions achieved high mapping accuracy on simulated Nanopore reads, while Winnowmap2 aligned more reads at mapQ10 or higher but sometimes overassigned mapping quality.
- 1,378,092 Winnowmap2 mapQ10 reads were all mapped by minimap2 v2.22, with 118 reads mapped differently and fewer than 0.01% of all reads affected.Of the 118 differences, 51 had multiple identical best alignments and were considered more likely Winnowmap2 errors.
- Most of the remaining 67 differing reads had multiple highly similar but non-identical alignments, with differences concentrated in highly repetitive regions.
- In simulated SV mapping, minimap2 v2.22 was as good as Winnowmap2, although a Sniffles threshold of 20 would miss five additional SVs.Those five represented 0.5% of simulated SVs, and the simulation omitted non-allelic gene conversions common in segmental duplications.
- For real contig-to-reference alignments involving INDELs longer than 1kb, v2.22 was more sensitive at comparable specificity.
- V2.22 had performance similar to older minimap2 versions and was consistently several times faster than Winnowmap2.The improvements added little computation because base alignment remained unchanged.