Sequence Alignment · Bioinformatics · Analysis Software · ComparisonSources read 17 October 2026
How DNA and protein sequences are lined up
Seven sequence aligners, three jobs
BLASTN, BLAT, LAST, BWA-MEM, Bowtie2, CLUSTALW and Clustal Omega are often listed together as “aligners”, but they answer three different questions. Three of them search for where one sequence resembles another; two place millions of sequencing reads on a reference genome; two line up many related sequences at once. Within each group the differences come down to what is indexed, how the first short matches (the seeds) are chosen, and how confidence is reported. This comparison reads each tool from its own paper, manual and source code, and ends in which tool fits which job.
What Happened
What is established about each tool, and how firmly Each item with its standing first, read in the tools’ own papers, manuals and source code on 17 October 2026.
- 3different jobs the seven tools do: local search, read mapping, multiple alignment
- 399 → 0.1chance hits per 500-base query on a 3-billion-base genome: one 14-mer seed against two 11-mers (Kent, 2002)
- 19 ntBWA-MEM’s default minimum seed length
- 190,000+sequences Clustal Omega aligned on one processor in a few hours, in its authors’ 2011 test
| Standing | What | Read at |
|---|---|---|
| Confirmed: current documentation or code | BLASTN builds its lookup table from the query’s words. The default task is megablast, with 28-base words; the blastn task uses 11, blastn-short 7, and dc-megablast 11 matching positions inside an 18-position template. | BLAST+ source code; Camacho et al. 2009 |
| Confirmed: current documentation or code | BLAT indexes the target genome’s non-overlapping 11-mers in memory and, for DNA, starts an alignment on two perfect 11-mer hits. Its default output (PSL) carries scores and identity, not an E-value. | Kent 2002; UCSC BLAT specification |
| Confirmed: current documentation or code | LAST indexes the reference as a suffix array and lengthens each seed until it occurs no more than m times there. It reports E-values and EG, the expected hits per billion query letters. | LAST documentation |
| Confirmed: current documentation or code | BWA-MEM seeds with supermaximal exact matches of at least 19 bases, re-seeds long ones, extends by banded dynamic programming and chooses between local and end-to-end alignment itself. It reports mapping quality (MAPQ). | Li 2013; bwa manual |
| Confirmed: current documentation or code | Bowtie2 looks up fixed seeds of 22 bases (20 in local mode) in an FM index of about 3.2 GB for the human genome, extends them with SIMD-accelerated dynamic programming, aligns end-to-end by default and reports MAPQ. | Bowtie2 manual; Langmead and Salzberg 2012 |
| Confirmed: the method paper | CLUSTALW aligns all pairs to get distances, builds a Neighbour-Joining guide tree, then aligns progressively along it with sequence weights, divergence-dependent matrices and position-specific gap penalties. | Thompson, Higgins and Gibson 1994 |
| Confirmed: the method paper | Clustal Omega builds its guide tree with mBed in O(N log N) time instead of O(N²) and aligns profile hidden Markov models with HHalign. | Sievers et al. 2011 |
| Developers’ own benchmark | BLAT “500 times faster” than the mRNA/DNA tools of 2002; adaptive seeds reaching 67% sensitivity in about 10 minutes against 500 for fixed seeds; BWA-MEM about six times as fast as Bowtie2 on 650-base reads; Clustal Omega’s BAliBASE score of 0.554 in 9 minutes. | Each tool’s own paper |
| Sources disagree | BLAT’s minimum DNA query length: the UCSC FAQ says 40 bases or more at 95% similarity; the UCSC Genome Browser User Guide says 25 bases, and that perfect matches of 32 bases, sometimes 22, are found. | UCSC BLAT FAQ; UCSC Genome Browser User Guide |
What is settled
How each tool indexes, seeds, extends and reports is stated in its own documentation, code or method paper.
What is the developers’ word
Every speed and accuracy figure comes from the tool’s own authors, on the hardware, data and rival tools of 1990 to 2013.
Timeline
Thirty-six years of aligners Papers dated by their journal issue or posting; releases by the projects’ own records.
- 1990Altschul and colleagues publish BLAST, which approximates alignments maximising a local similarity score and attaches statistical significance to them.
- 11 Nov 1994Thompson, Higgins and Gibson publish CLUSTAL W: sequence weighting, matrices varied by divergence and position-specific gap penalties.
- 1 Sep 1997Gapped BLAST and PSI-BLAST: a new trigger for extending word hits and a gapped-alignment heuristic, about three times the speed of the original for proteins.
- Apr 2002Kent publishes BLAT, which indexes the genome rather than the query and stitches exons into gene-level alignments.
- 15 Dec 2009Camacho and colleagues describe BLAST+, whose “tasks” set word size and scoring for each kind of search.
- Mar 2011Kiełbasa, Frith and colleagues publish adaptive seeds and LAST: seeds chosen by rarity, so run time grows linearly rather than quadratically with sequence length.
- 11 Oct 2011Sievers and colleagues publish Clustal Omega: mBed guide trees and HHalign, tested on up to 50,000 sequences.
- 4 Mar 2012Langmead and Salzberg publish Bowtie 2: FM-index seeds, then SIMD-accelerated gapped extension.
- 2013Shrestha and Frith publish a probabilistic method for paired-end reads, implemented in LAST.
- 26 May 2013Li posts version 2 of the BWA-MEM preprint: SMEM seeds, re-seeding, chaining and Z-drop extension, for 70 bases to a few megabases.
- 2015Frith and Kawaguchi show split alignment of genomes that gains accuracy by not masking repeats.
- 2017Hamada, Frith and colleagues publish last-train, which learns substitution and gap scores from the data being aligned.
- 2021Sievers and Higgins’ protocol chapter: Clustal Omega handles many tens of thousands of DNA/RNA or protein sequences.
- 22 Mar 2025BWA 0.7.19 released (GitHub, UTC date).
- Jul 2025BLAST+ 2.17.0: the release NCBI’s download directory lists as latest, its files dated 1 July.
- 16 Feb 2026Bowtie2 2.5.5 released.
- 3 Aug 2026LAST tagged as version 1654, the latest in its repository on the day these sources were read.
What cannot be dated. BLAT’s current release, Clustal Omega’s current release and CLUSTALW’s last release: none of the sources read gives a date, and Clustal Omega’s own site refused the request. The tools’ defaults on this page are those of the current documentation and code, not necessarily those of the papers.
The Argument
Three jobs, and inside each one the index and the seed decide Each source describes only its own tool; the grouping is this article’s reading of them.
Search
Where does this sequence resemble that one?
- Local alignment: only the best-matching stretches are reported.
- BLASTN indexes the query and scans the database; BLAT and LAST index the database and scan the query.
Map reads
Where in this genome did each read come from?
- The reference is indexed once with an FM index; millions of reads stream through it.
- Pairs are handled, and the answer comes with a mapping quality.
Align many
How do all these related sequences line up together?
- No index and no seeds: distances give a guide tree, and sequences join along it.
- Every position of every sequence is placed, end to end.
| Tool | What is indexed | Seed |
|---|---|---|
| BLASTN | The query’s words, in a lookup table; the database is scanned | Exact words: 28 bases (megablast, default), 11 (blastn), 7 (blastn-short); spaced in dc-megablast |
| BLAT | Non-overlapping 11-mers of the target genome, held in memory | Two perfect 11-mer hits on one diagonal (DNA) |
| LAST | A suffix array of the reference | Adaptive: each seed lengthened until it occurs at most m times |
| BWA-MEM | FM index of the reference | Supermaximal exact matches, at least 19 bases; long ones re-seeded |
| Bowtie2 | FM index of the reference | Fixed exact seeds of 22 bases (20 in local mode), spaced by read length |
| CLUSTALW | Nothing; all pairs are aligned for distances | None; a Neighbour-Joining guide tree instead |
| Clustal Omega | Nothing; each sequence becomes a short vector of distances (mBed) | None; a clustered guide tree instead |
Why the seed rule matters. A seed is a short exact match found quickly through an index; only seeded regions get the expensive alignment step, so the seed rule sets both the speed and what is missed. Kent worked the arithmetic for BLAT in 2002: to find 99% of 100-base regions at 95% identity, a single perfect seed must be 14 bases or shorter, and against a 3-billion-base genome that passes 399 chance regions per 500-base query to alignment; two perfect 11-mers pass 0.1. LAST’s answer is different: rather than fixing the length, it lengthens each seed until it is rare. Its developers measured 777 million matches with adaptive seeds between the human and mouse X chromosomes, against 22 billion for fixed 13-base seeds, and argued that adaptive seeds give the sensitivity of a 13-base seed at the run time of a 32-base one.
- 399chance regions, one 14-mer seed
- 0.1chance regions, two 11-mer seeds
A reading
Reading. The first decision is the job, not the tool. Within search, the trade-off between speed and sensitivity is set mostly by the seed; within read mapping, by how seeds are found and extended in an FM index; within multiple alignment, by how the guide tree is built.
What Others Add
Extension, scoring and what the numbers mean From each tool’s current manual or code where one exists, otherwise its method paper.
| Tool | Extension | Default scoring | Confidence reported | Typical input |
|---|---|---|---|---|
| BLASTN | Ungapped, then gapped X-drop | blastn +2/−3, gaps 5 + 2 per base; megablast +1/−2, linear gaps; low-complexity filter on | E-value and bit score (default cut-off 10) | A query against a database |
| BLAT | Align the clumped regions, stitch the blocks, revisit small exons and splice sites | Matches minus mismatches minus a gap penalty; minimum score 30, minimum identity 90% | Score and identity; no E-value in the default output | mRNA or DNA up to 200,000 bases |
| LAST | Gapless, then gapped X-drop from a core | A matrix or match/mismatch scores, or scores learnt by last-train; gaps a + b·k | E and EG; by default, matches expected by chance at most once per million query letters | Reads to whole genomes |
| BWA-MEM | Banded affine-gap dynamic programming with Z-drop; mate rescue | Match 1, mismatch 4, gap open 6, extension 1, clipping 5 | Mapping quality (MAPQ) | 70 bases to a few megabases |
| Bowtie2 | SIMD-accelerated dynamic programming | End-to-end: mismatch up to 6 by base quality, read gap 5 + 3 per base; local mode adds a match bonus of 2 | Mapping quality (MAPQ) | Reads of about 50 to a few hundred bases |
| CLUSTALW | Full profile-to-profile dynamic programming at each tree node | Matrices varied by divergence; residue- and position-specific gap penalties | None per alignment | Small sets of related sequences |
| Clustal Omega | Profile hidden Markov model against profile HMM (HHalign) at each node | Profile HMMs; an external profile can guide the alignment | None per alignment | Tens of thousands of DNA, RNA or protein sequences |
An E-value and a MAPQ answer different questions. BLASTN and LAST report how often a score this good would arise between random sequences of the same lengths, which suits a search where any hit might be chance. BWA-MEM and Bowtie2 report a Phred-scaled estimate of how likely the read has been placed in the wrong spot, which suits a read assumed to come from the genome that may fit several places. BLAT’s default output gives match counts and identity, suited to near-identical placement. The multiple aligners report no per-alignment figure at all: their quality is judged on benchmark sets, which their own authors call “heavily biased toward single domain globular proteins”. None of these converts into another.
Repeats, memory and long reads. BLAT leaves overly frequent 11-mers out of its index; its server index takes about 1 GB for unmasked DNA. LAST’s adaptive seeds let genomes be compared with little or no repeat masking, though its authors still advise masking simple repeats; its split alignment of genomes “gains accuracy by not masking repeats”. Its index grows with the seeding scheme: the most sensitive, MAM8, needs about 50 GB for a mammal genome, while the RY schemes cut time and memory by seeding at only some positions. For long reads with many insertion and deletion errors, LAST’s own cookbook trains parameters with last-train and then keeps one best alignment per part of each read; last-train’s authors report better accuracy for reads from three sequencing platforms and less reference bias for nanopore reads. BWA-MEM covers 70 bases to megabases; its author found seeding the bottleneck for short reads and banded alignment for long ones. Bowtie2’s FM index of the human genome uses about 3.2 GB.
- ~1 GBBLAT server index, unmasked DNA
- ~3.2 GBBowtie2 index, human genome
- ~50 GBLAST with MAM8 seeds, mammal genome
Benchmarks of their day
BLAT’s 500-fold speed-up was against tools of 2002; Bowtie2’s 2012 memory comparison was with the original BWA, a year before BWA-MEM; BWA-MEM’s sixfold lead over Bowtie2 was on 650-base reads, and the two were similar on 100-base reads. Each figure is its authors’ own.
Clustal Omega against the consistency methods
On BAliBASE its authors measured 0.554 of columns correct in 9 minutes; consistency-based programs such as Probcons and T-Coffee reached 0.55 to 0.61 in 25 minutes to 22 hours, and Clustal W was the weakest program tested.
Conclusion
Which tool suits which job One card per tool: who it is for, what it is best at, and the trade-off.
BLASTN
Finding related sequences in a database, with a statistic attached
- The default task is the near-identical one; divergent DNA needs
blastnordc-megablastchosen on purpose. - Trade-off: scanning the database for every query is slow at the scale of whole genomes or millions of reads.
- The default task is the near-identical one; divergent DNA needs
BLAT
Placing an mRNA, cDNA or sequence on its own species’ genome
- Designed for DNA at 95% similarity or more, and protein at 80% or more.
- Trade-off: it misses divergent matches, short sequences that span a splice site, and gives no E-value by default.
LAST
Comparing whole genomes, or reads whose error profile is unknown
- Adaptive seeds, split alignment, E-values, and DNA-to-protein as well as DNA-to-DNA.
- Trade-off: memory grows with the seeding scheme, and pairing is a separate step rather than built in.
BWA-MEM
Mapping reads or contigs from 70 bases to megabases
- Paired reads are rescued and scored using the insert-size distribution.
- Trade-off: it answers where a read belongs, not how significant a match is; its speed lead in its author’s test was on long reads.
Bowtie2
Mapping short reads to a large genome
- Concordant, discordant and mixed pairing, with quality-aware scoring.
- Trade-off: built for reads up to a few hundred bases; on 650-base reads BWA-MEM’s author measured it about six times slower than his tool.
CLUSTALW
Reproducing an older analysis, or teaching progressive alignment
- Its method and its gap rules are set out in full in its 1994 paper.
- Trade-off: weakest on its successor’s benchmark, O(N²) to build the tree, and unreliable when every pair is below about 25–30% identity.
Clustal Omega
Multiple alignment at any size, from a handful to tens of thousands
- Handles DNA, RNA and protein; iteration helps most near 1,000 sequences and above.
- Trade-off: on BAliBASE, whose test sets are mostly small, the best consistency methods scored somewhat higher (up to 0.61 against 0.554), at far greater run time.
Hold loosely
Defaults change between versions; these are the defaults of the documentation and code read on 17 October 2026. Every speed figure is the developers’ own, from 1990 to 2013, and no independent benchmark of all seven was found.
- Reading. Short reads onto a reference genome: BWA-MEM or Bowtie2; both need the reference indexed first.
- Reading. Longer reads or assembled contigs: BWA-MEM; for long reads with many indel errors, LAST with
last-train. - Reading. An mRNA or cDNA on its own species’ genome at high identity: BLAT.
- Reading. Divergent homologues that need an E-value: BLASTN with the
blastnordc-megablasttask. - Reading. Whole genomes, especially repeat-rich ones: LAST.
- Reading. Multiple alignment: Clustal Omega; CLUSTALW where an older result must be reproduced.
Quick notes
- W. J. Kent, BLAT (Genome Research, 2002, PMC)
- S. M. Kiełbasa and colleagues, adaptive seeds (Genome Research, 2011, PMC)
- LAST documentation (lastal, E-values, seeding schemes, lastdb, last-train, cookbook) and repository tags, GitLab
- M. Hamada and colleagues, last-train (Bioinformatics, 2017, PubMed)
- A. M. Shrestha and M. C. Frith, paired-end reads (Bioinformatics, 2013, PubMed)
- M. C. Frith and R. Kawaguchi, split alignment (Genome Biology, 2015, PMC)
- H. Li, BWA-MEM (arXiv, 2013), bwa manual page and release 0.7.19, GitHub
- Bowtie 2 manual and NEWS, GitHub
- B. Langmead and S. L. Salzberg, Bowtie 2 (Nature Methods, 2012, PMC)
- S. F. Altschul and colleagues, BLAST (Journal of Molecular Biology, 1990, PubMed) and gapped BLAST (Nucleic Acids Research, 1997)
- C. Camacho and colleagues, BLAST+ (BMC Bioinformatics, 2009, PMC)
- NCBI C++ Toolkit BLAST source code, GitHub
- NCBI BLAST+ download directory
- UCSC Genome Browser, BLAT FAQ, BLAT specification and User Guide
- J. D. Thompson, D. G. Higgins and T. J. Gibson, CLUSTAL W (Nucleic Acids Research, 1994)
- F. Sievers and colleagues, Clustal Omega (Molecular Systems Biology, 2011, PMC)
- F. Sievers and D. G. Higgins, the Clustal Omega package (Methods in Molecular Biology, 2021, PubMed). Read 17 October 2026.
Reliable
Worth believing
Sharing a subject
- Six single-cell analysis platforms, comparedAnalysis Software · Bioinformatics
- The Filters That Decide What a Single-Cell Experiment SawBioinformatics
- Small Variants in the Long-Read EraBioinformatics
Sharing a source domain
- Avalanche's RWA Boom, and Why AVAX Need Not Followgithub.com · githubusercontent.com
- Five Names for One Stack, and No Agreed Orderarxiv.org · github.com
- Enough, or Less? Context Engineering Said Both in One Weekarxiv.org · github.com
- It Began as Tuning an Agent. By March It Meant Everything but the Model.arxiv.org · github.com
- One Term, Coined Twice, Two Weeks Apart. The Joke Won.arxiv.org · github.com
- Dorado v4 to v6 — What Changed, How Much, and Whether It Was Goodgithub.com · nih.gov
- Forty-Three Years Was Last Month’s Pricegithub.com · githubusercontent.com
- Meal Sequencing — Vegetables Before Starchnih.gov
- Dark Marks on Clothing — Identifying and Removing Themnih.gov
- The AISI Cyber Evaluation of Claude Mythos Previewarxiv.org
- The Claude Code Quality Postmortem — Three Unrelated Changesgithub.com
- The Rise of Biomedical Foundation Models — 220 Models Surveyednih.gov
- Local LLMs and Buying a Mac — A Technical Reportgithub.com
- 9 Billion Requests a Day, Moved Without Rebuilding a Filegithub.com
- Grok Build Shipped Whole Repos to xAI's Cloudgithub.com
- Cryptographic Weakness Discovery — Reading Anthropic Against Its Own Papersarxiv.org
- Swap Across Five Platforms — Compress First, Spill Latergithubusercontent.com
- AI Reasoning Traces: Encrypted, Not Hiddenarxiv.org
- The Tricks Moved Into the Model. The Writing Stayed.arxiv.org
- Everyone Agrees You Should Write Loops. Nobody Agrees What One Is.arxiv.org
- The Harness Ships, the Governance Doesn’tarxiv.org
- The Polyamory Words, and the One Nobody Can Datenih.gov
- The Electron, to Thirteen Digitsarxiv.org
- Four Lepton Anomalies, One Survivorarxiv.org
- Six Leptons, a Word Made for Twoarxiv.org
- Nineteen Numbers, Three Forces of Fourarxiv.org
- Three Cuts Into the Uncuttablearxiv.org
- Shunting Bulk Reads to a Cheaper Model — What the 90% Actually Cutsgithub.com
- The Dying, Not the Bereavednih.gov
- The Colour Is the Vitaminnih.gov
- The Switch Nobody Has Foundnih.gov
- The Sun Does Not Kill Themnih.gov
- Dairy in East and Southeast Asianih.gov
- Six single-cell analysis platforms, comparedgithub.com
- Small Variants in the Long-Read Eranih.gov
- REST, SOAP, gRPC, GraphQL and MCP: pick by who callsgithub.com