bonefish

A fast aligner for Oxford Nanopore and PacBio HiFi reads, written in Rust. bonefish maps ONT genomic DNA, direct RNA and cDNA, and PacBio HiFi genomic DNA, to a reference genome. It writes SAM, PAF or a sorted, indexed BAM in one step.

• Accurate. On whole human genomes (GIAB HG002 and HG005), SNP, indel and structural-variant calls match minimap2's. Methylation calls agree at r = 0.999.
• Fast.
  • Aligner alone: 2.6-8x faster than minimap2 on DNA, 2.6x on direct RNA, 5.4x on cDNA and 2.3-2.8x for transcriptome quantification. On PacBio HiFi it needs 3.9x less CPU time than minimap2 map-hifi.
  • Reads to a sorted BAM: about 3x faster than minimap2 | samtools sort.
• Drop-in. It reads dorado's BAM output, keeping methylation and other tags, and accepts minimap2's options and preset names.

Install

Prebuilt binaries: download from the latest release. They are available for Linux x86_64 and arm64 (glibc 2.28+: RHEL/Rocky 8+, Ubuntu 20.04+, Debian 10+) and macOS on Apple silicon. Unpack and run ./bonefish.

From source, with a recent stable Rust toolchain:

cargo build --release --features jemalloc

Container:

docker build -t bonefish https://github.com/DylanCosto/bonefish.git

Quick start

# DNA: dorado BAM in, sorted BAM + index out (-y keeps methylation tags)
bonefish -t 16 -y -o sample.bam ref.fa reads.bam

# Direct RNA and cDNA, spliced to the genome
bonefish -t 16 -x rna  -y -o rna.bam  ref.fa reads.bam
bonefish -t 16 -x cdna -y -o cdna.bam ref.fa reads.bam

# FASTQ in, SAM out
bonefish -t 16 -a ref.fa reads.fq.gz > sample.sam

For repeated runs, build an index once: bonefish -d ref.raw --idx-raw ref.fa. It loads in under a second; add -x rna or -x cdna for a spliced index. New to bonefish? The tutorial goes from ONT's public demo data to a sorted BAM and methylation calls in a few minutes.

preset	for
dna (default; alias ont)	genomic DNA: R10.4.1 or R9.4.1, sup or hac, ultra-long (minimap2 lr:hq)
hifi	PacBio HiFi genomic DNA (minimap2 map-hifi)
rna	direct RNA (minimap2 splice -uf -k14)
cdna	cDNA (minimap2 splice)
tx	RNA or cDNA to a transcriptome, for quantification with oarfish or NanoCount (minimap2 map-ont -N 100)

Input formats, output options, memory use and CPU support are covered in docs/usage.md. To replace minimap2 in an existing pipeline, including ONT's wf-human-variation, see docs/pipelines.md.

Speed and accuracy

All runs below used 16 threads and compare against minimap2 2.31 on the same reads.

Whole human genomes, reads to a sorted, indexed BAM (v0.1.5; HG005 timed with v0.1.1). Each F1 cell gives minimap2 / bonefish:

sample (R10.4.1 sup)	minimap2 + samtools	bonefish	SNP F1	indel F1	SV F1
HG002, 27x	115 min	38 min	0.9971 / 0.9971	0.8737 / 0.8737	0.8055 / 0.8053
HG005, 28x	115 min	35 min	0.9975 / 0.9975	0.9395 / 0.9395	–

Small variants were called with Clair3 against GIAB v4.2.1, and structural variants with Sniffles2 against GIAB v5.0q.

Aligner alone:

dataset	minimap2	bonefish	speedup
DNA, HG002 ultra-long (3.4 Gbp)	542 s	65 s	8.3x
DNA, HG001 genome-wide sample (2.7 Gbp)	202 s	44 s	4.6x
DNA, HG002 R9.4.1 chr20 (4.6 Gbp)	412 s	103 s	4.0x
cDNA, HG002 PCS114 (809k reads)	1,223 s	227 s	5.4x
direct RNA, HG002 RNA004 (902k reads)	456 s	177 s	2.6x
direct RNA to the transcriptome (-x tx, 200k reads)	106 s	37 s	2.8x

On RNA, 99.5-99.9% of primary alignments are identical to minimap2's, and junction accuracy against GENCODE is the same. Transcript quantification from -x tx alignments agrees with minimap2's at gene level at r = 0.9995-0.99997.

bonefish is also faster than dorado's aligner (4.3x), mm2plus (5.1x), mm2-fast (14x) and Winnowmap2 (38x).

• Datasets and comparisons: docs/benchmarks.md has the datasets, commands, every comparison, and the known differences from minimap2.
• Raw evidence: docs/validation indexes the reports behind each number.
• Other read types: the reports cover metagenomes, duplex reads, other references (T2T-CHM13, ALT contigs), simulated reads with known origin, medically relevant genes, tandem repeats and output checks.

How it works

bonefish uses minimap2's seed-chain-align method with the same scoring, so its output stays close to minimap2's. Most of the speed comes from doing less dynamic-programming (DP) alignment:

• DP bands: narrower, adaptive bands, widened only where an indel needs it.
• Gap filling: gaps are first tried without indels, and long gaps are split at unique matches.
• Introns: they are aligned against just the reference next to the chain's anchors. A 10-mer search looks for exons that have no anchor.
• cDNA strands: each chain is aligned once for both transcript strands.

The rest comes from engineering:

• SIMD kernels: hand-written AVX-512, AVX2 and NEON kernels for chaining, DP, seeding and sorting, chosen at run time.
• I/O and sorting: multithreaded I/O, an index that loads in under a second, and a built-in BAM sorter.

Limitations

• PacBio support covers HiFi genomic DNA (-x hifi). Iso-Seq/Kinnex RNA has not been validated, and there are no short-read or assembly presets.
• The RNA presets are new (v0.2.0). They are validated against minimap2, GENCODE and simulated reads, but not yet with downstream isoform or expression tools.
• Output is SAM, PAF or sorted BAM; there is no CRAM output.
• On low-accuracy reads (about 93% identity), minimap2's map-ont maps about 1% more reads.

Library

bonefish-core (imported as bonefish) exposes the aligner as a Rust library: bonefish::Aligner with Preset::Dna, Preset::Hifi, Preset::Rna and Preset::Cdna. See bonefish-core/examples/.

Contributing

Bug reports and pull requests are welcome; see CONTRIBUTING.md. Changes in each release are listed in CHANGELOG.md.

Credits and license

bonefish is a fork of rammap, a pure-Rust reimplementation of minimap2 by Heng Li. The index format, SIMD DP kernels and most of the mapping pipeline come from rammap, and the alignment algorithms are minimap2's. bonefish is MIT licensed (see LICENSE), as both of those projects are.