LASTZ   Release 1.04.15, built August 27, 2021

• Introduction
• Availability
• Installation
• Build Options
• Overview of Processing Stages and Terminology
• Examples
  • Comparing a Human Chromosome and a Chicken Chromosome
  • Aligning Shotgun Reads to a Human Chromosome
  • Seeds, HSPs, Gapped Alignments, Chaining
  • Aligning a Sequence With Itself
• Command-line Syntax
  • Where to Look
  • Scoring
  • Indexing
  • Seeding
  • HSPs (Gap-free Extension)
  • Chaining
  • Gapped Extension
  • Back-end Filtering
  • Interpolation
  • Output
  • Housekeeping
  • Shortcuts for Yasra
  • Help
  • Sequence Specifiers
• Processing Stages in Detail
  • Target Sequence Input
  • Scoring Inference
  • Indexing Target Seed Words
  • Seeding
  • Gap-free Extension
  • HSP Chaining
  • Gapped Extension
  • Back-end Filtering
  • Interpolation
  • Alignment Output
• File Formats
  • FASTA
  • FASTQ
  • Nib
  • 2Bit
  • Quantum DNA
  • Quantum Code File
  • Sequence Name File
  • Sequence Masking File
  • Sequence Masking File, Three Fields
  • Scoring File
  • Inference Control File
  • HSX (Hashed Sequence Index)
  • Target Capsule File
  • Alignment Chores File
  • Segment File
  • LAV
  • AXT
  • MAF
  • SAM
  • CIGAR
  • BLASTN
  • PAF
  • PAF:wfmash
  • Differences
  • R Dotplot
  • Human-Readable Text
  • General Output
  • Other Output
• Advanced Topics
  • Aligning to Whole Genomes
  • Adjacent Indels
  • Interval Coordinates
  • Non-ACGT Characters, Splicing, and Separation
  • Sequence Name Mangling
  • Seed Patterns
  • Any-or-None Alignment
  • Y-drop Mismatch Shadow
  • Shingle Overlap
  • Using Target Capsule Files
  • Inferring Score Sets
  • Dynamic Programming Matrix
  • Filtering With Shell Commands
  • Self-Masking a Sequence
  • Aligning Many Subintervals
• Differences from BLASTZ
• Change History
  • Most Recent Changes
• References
• Acknowledgments

Introduction

This document describes installation and usage of the LASTZ sequence alignment program. LASTZ is a drop-in replacement for BLASTZ, and is backward compatible with BLASTZ’s command-line syntax. That is, it supports all of BLASTZ’s options but also has additional ones, and may produce slightly different alignment results.

Availability

LASTZ is available from github at https://github.com/lastz/lastz.

A packed archive containing source code for LASTZ is available from the Miller Lab at Penn State. Note: 1.04.15 is the last release that will be available via the Miller Lab website.

Installation

If you have received the distribution as a packed archive, unpack it by whatever means are appropriate for your computer. The result should be a directory <somepath>/lastz‑distrib‑X.XX.XX that contains a src subdirectory (and some others). You may find it convenient to remove the revision number (‑X.XX.XX) from the directory name.

Before building or installing any of the programs, you will need to tell the installer where to put the executable, either by setting the shell variable $LASTZ_INSTALL, or by editing the make‑include.mak file to set the definition of installDir. Also, be sure to add the directory you choose to your $PATH.

Then to build the LASTZ executable, enter the following commands from bash or a similar command-line shell (Solaris users should substitute gmake for make). This will build two executables (lastz and lastz_D) and copy them into your installDir.

    cd <somepath>/lastz-distrib-X.XX.XX/src
    make
    make install

The build process should not report any warnings or errors. Because of this, the Makefile is set up so that warnings are considered errors and will stop the build. If you encounter this situation, you can use Makefile.warnings instead:

    make -f Makefile.warnings

A simple self test is included so you can test whether the build succeeded. To run it, enter the following command:

    make test

    make: *** [test] Error 1

Build Options

An additional executable (lastz_32) can be built, to handle genomes larger than 2 gigabases. For details, see the section on aligning to whole genomes.

Any executable can be built to allow adjacent indels (by default, these are not allowed). For details, see the section on adjacent indels.

Overview of Processing Stages and Terminology

LASTZ is designed to preprocess one sequence or set of sequences (which we collectively call the target) and then align several query sequences to it. The general flow of the program is like a pipeline: the output of one stage is the input to the next. The user can choose to skip most stages via command-line options; any stages that are skipped pass their input along to the next stage unchanged. Two of the stages, scoring inference and interpolation, are special in that they perform a miniature version of the pipeline within them.

Note that the following discussion is a generalization, intended to describe the basic idea of LASTZ’s operation. There are many exceptions that depend on the particular options specified.

The stages are:

• read sequence(s) from target file
• infer scoring parameters (if necessary)
• build a seed word position table for the target
• for each sequence in query file:
  • find seeds
  • perform gap-free extension of seeds to HSPs
  • chain HSPs
  • perform gapped extension of HSPs to local alignments
  • filter alignments according to specified criteria
  • interpolate between alignments
  • print output
  • repeat with the reverse complement of the query

The usual flow is as follows (though most of these steps are optional, and some settings like ‑‑anyornone may affect the processing order). We first read the target sequence(s) into memory, and use that to build a seed word position table that will allow us to quickly map any word in the target to all of the positions where it appears. (For the purposes of this discussion you can think of a word as a 12-mer of DNA.) Then we read each query sequence in turn, processing them more or less independently. We examine the word starting at each base in the query and use the position table to find matches, called seeds, in the target. The seeds are extended to longer matches called HSPs (high-scoring segment pairs) and filtered based on score. The HSPs are chained into the highest-scoring set of syntenic alignments, and then reduced to single locations called anchors. The anchors are then extended to local alignments (which may contain gaps) and again filtered by score, followed by back-end filtering to discard alignment blocks that do not meet specified criteria for certain traits. We then interpolate, repeating the entire process at a higher sensitivity in the holes between the alignment blocks. And finally, we write out the alignment information to a file. Then these steps are repeated with the reverse complement of the query sequence, before moving on to the next sequence in the query file.

The scoring inference stage is not usually performed. Typically it is used only when sequences for new species are acquired, to create scoring files for subsequent alignments of those species.

Examples

For those eager to try it out, here are some illustrative examples to get you started. Detailed reference material begins with the next section.

Comparing a Human Chromosome and a Chicken Chromosome

It is often adequate to use a lower sensitivity level than is achieved with LASTZ’s defaults. For example, to compare two complete chromosomes, even for species as distant as human and chicken, the alignment landscape is evident even at very low sensitivity settings. This can speed up the alignment process considerably.

This example compares human chromosome 4 to chicken chromosome 4. These sequences can be found in the downloads section of the UCSC Genome Browser, and are 191 and 94 megabases long, respectively. To run a quick low-sensitivity alignment of these sequences, use a command like this:

    lastz hg18.chr4.fa galGal3.chr4.fa \
      --notransition --step=20 --nogapped \
      --format=maf > hg18_4_vs_galGal3_4.maf

This runs in about two and a half minutes on a 2-GHz workstation, requiring only 400 Mb of RAM. Figure 1(a) shows the results, plotted using the ‑‑format=rdotplot output option and the R statistical package. (When in MAF format, LASTZ output can be browsed with the GMAJ interactive viewer for multiple alignments, available from the Miller Lab at Penn State.)

Using ‑‑notransition lowers seeding sensitivity and reduces runtime (by a factor of about 10 in this case). ‑‑step=20 also lowers seeding sensitivity, reducing runtime and also reducing memory consumption (by a factor of about 3.3 in this case). ‑‑nogapped eliminates the computation of gapped alignments. The complete alignment process using default settings (shown in Figure 1(b)) uses 1.3 Gb of RAM and takes 4.5 hours on a machine running at 2.83 GHz.

lastz \
  hg18.chr4.fa galGal3.chr4.fa \
  --notransition --step=20 \
  --nogapped

lastz \
  hg18.chr4.fa galGal3.chr4.fa

Aligning Shotgun Reads to a Human Chromosome

Short read mapping for close species requires parameters very different from LASTZ’s defaults. This example compares a simulated set of primate shotgun reads to human chromosome 21. The chromosome can be found in the downloads section of the UCSC Genome Browser (it is about 47 megabases). Ten thousand simulated reads were generated by extracting 60-bp intervals from chimp chr21, subjecting them to mild mutation (including short gaps), and then truncating them to 50 bp (these are included in the LASTZ distribution, in test_data/fake_chimp_reads.2bit).

To see where these reads map onto the human chromosome, use this command:

    lastz hg18.chr21.fa[unmask] fake_chimp_reads.2bit \
      --step=10 --seed=match12 --notransition --exact=20 --noytrim \
      --match=1,5 --ambiguous=n \
      --filter=coverage:90 --filter=identity:95 \
      --format=general:name1,start1,length1,name2,strand2 \
      > hg18_21_vs_reads.dat

Attaching [unmask] to the chromosome filename instructs LASTZ to ignore masking information and treat repeats the same as any other part of the chromosome, in order to accurately assess the uniqueness of the read mappings. Since we know the two species are close, we want to reduce sensitivity. Using ‑‑step=10, we will only be looking for seeds at every 10th base. Instead of the default seed pattern, we use ‑‑seed=match12 and ‑‑notransition so our seeds will be exact matches of 12 bases. Instead of the default x-drop extension method we use ‑‑exact=20 so that a 20-base exact match is required to qualify as an HSP. Because we are aligning short reads, we specify ‑‑noytrim so the alignment ends will not be trimmed back to the highest scoring locations during gapped extension.

We replace the default score set, which is for more distant species, with the stricter ‑‑match=1,5. This scores matching bases as +1 and mismatches as −5. We also use ‑‑ambiguous=n so that Ns will be scored appropriately. We are only interested in alignments that involve nearly an entire read, and since the species are close we don't want alignments with low identity; therefore we use ‑‑filter=coverage:90 and ‑‑filter=identity:95.

For output, we are only interested in where the reads align, so we use the ‑‑format=general option and specify that we want the position on the chromosome (name1, start1, length1) and the read name and orientation (name2, strand2). This creates a tab-delimited output file with one line per alignment block, a format that is well-suited for downstream processing by other programs. For example, to count the number of different reads we've mapped, we can run this Unix shell command:

    cat hg18_21_vs_reads.dat | grep -v "#" | awk '{print $4}' | sort -u | wc

Seeds, HSPs, Gapped Alignments, Chaining

This example demonstrates the primary alignment processing stages, using the α-globin regions of cow and human. This data is included in the LASTZ distribution in test_data/aglobin.2bit, and consists of a 70K bp segment of human DNA and a 66K bp segment of cow DNA. We will follow this example through the major stages of seeding, gap-free extension, chaining, and gapped extension.

Figure 2(a) shows the result of default seeding on a small window (3K bp) in the middle of these segments. Seeds are short near-matches; in this case each seed is 19 bp and could have as many as 8 mismatches (12-of-19 with one transition). There are 338 seeds in this window, but regions where there are many seeds are indistinguishable from line segments.

Figure 2(b) shows high-scoring segment pairs, the result of gap-free extension of the seeds. There are 11 HSPs (only 10 are apparent in the figure, but one of those is split by a 1-bp shift to the next diagonal). Note that many seeds were discarded because their extensions were low scoring or overlapped.

Figure 2(c) shows the local alignment blocks resulting from gapped extension of the HSPs. There are four alignment blocks.

Then we zoom out and show the results for the full sequences; the red box indicates the small region shown in the earlier figures. Figure 2(d) shows the HSPs, 2(e) shows the gapped alignment blocks, and 2(f) illustrates how chaining reduces the alignment blocks to a single syntenic line (or two lines, if there were matches on both strands). Note that one can already tell quite a bit about how the sequences align just from looking at the HSPs.

lastz \
  aglobin.2bit/human[34000..37000] \
  aglobin.2bit/cow[35000..38000] \
  --nogfextend --nochain --nogapped

lastz \
  aglobin.2bit/human[34000..37000] \
  aglobin.2bit/cow[35000..38000] \
  --gfextend --nochain --nogapped

lastz \
  aglobin.2bit/human[34000..37000] \
  aglobin.2bit/cow[35000..38000] \
  --gfextend --nochain --gapped

lastz \
  aglobin.2bit/human \
  aglobin.2bit/cow \
  --gfextend --nochain --nogapped

lastz \
  aglobin.2bit/human \
  aglobin.2bit/cow \
  --gfextend --nochain --gapped

lastz \
  aglobin.2bit/human \
  aglobin.2bit/cow \
  --gfextend --chain --gapped

Aligning a Sequence With Itself

When a sequence is aligned to itself, the full result will contain mirror-image copies of each alignment block. It is computationally wasteful to process both copies. LASTZ can handle this situation in four different ways.

1. Simply give LASTZ the same sequence for both the target and query. In this case, LASTZ does not know that it is aligning a sequence to itself, and performs the full computation on both copies (Figure 3(a)).

2. Specify the ‑‑notrivial option. This performs the full computation on both copies, but doesn't report the trivial self-alignment block along the main diagonal (Figure 3(b)).

3. Specify the ‑‑self option in place of the query sequence. LASTZ will save work by computing with only one block of each mirror-image pair, though it still reports both copies in the output by reconstructing the second copy from the first. It also invokes ‑‑notrivial automatically to omit the trivial self-alignment block along the main diagonal. This gives the same output as the previous method, but runs faster (Figure 3(c)).

4. Specify ‑‑self in place of the query, and also add the ‑‑nomirror option. In this case LASTZ reports only one copy of each mirror-image pair, as well as omitting the trivial block (Figure 3(d)).

In the following figure, we suppose we have a sequence with repeated motifs, in the order α1 β1 γ1 β2 δ1 α2 δ2′ γ2. That is, α1 and α2 are ancient duplications, as are β1 and β2, and γ1 and γ2.  δ2′ is an inversion, a reverse-complement duplicate of δ1.

lastz target target

lastz target target --notrivial

lastz target --self

lastz target --self --nomirror

Command-line Syntax

If you are familiar with BLASTZ, you can run LASTZ the same way you ran BLASTZ, with the same options and input files. In addition to this BLASTZ compatibility, LASTZ provides other options.

The general format of the LASTZ command line is

    lastz <target> [<query>] [<options>]

The angle brackets <> indicate meta-syntactic variables that should be replaced with your values, while the square ones [] indicate elements that are optional. Spaces separate fields on the command line; a field that needs to contain a space (e.g. within a file name) must be enclosed in double quotes "". Elements can appear in any order, the only constraint being that, if present, the <query> must appear after the <target>. Output is generally written to stdout, unless specified otherwise for a particular option.

The <target> and <query> are usually just the names of files containing the sequences to be aligned, in either FASTA, Nib, or 2Bit format. However they can be HSX index files that refer to the sequences indirectly, and they also can specify pre-processing actions such as selecting a subsequence from the file (see Sequence Specifiers for details). With certain options such as ‑‑self the <query> is not needed; otherwise if it is left unspecified the query sequences are read from stdin (though this does not work with random-access formats like 2Bit). As a special case, the <target> is omitted when the ‑‑targetcapsule option is used, since the target sequence is embedded within the capsule file.

For options, the general format is ‑‑<keyword> or ‑‑<keyword>=<value>, but for BLASTZ compatibility some options also have an alternative syntax <letter>=<number>. (Be careful when copying options from the tables below, as some of the hyphens here are special characters to avoid awkward line wrapping in certain web browsers. If you have trouble, replace the pasted hyphens with ordinary typed ones on your command line.)

Please understand that LASTZ is a complex program and its options are not all independent, i.e., some options are not valid in combination with certain others. It would be difficult and cumbersome to attempt to list every possible conflict here; instead we just mention some of the major ones. If you are not sure about a particular combination, go ahead and try it — LASTZ will tell you if it’s not allowed.

Running the command lastz without any arguments prints a help message with the most commonly used options, while running

    lastz --help

Where to Look

Scoring

These are fundamental parameters for alignment scoring, used in several of the stages.

Indexing

Seeding

Finding HSPs (Gap-free Extension)

Chaining

Gapped Extension

Back-end Filtering

Interpolation

Output

    # lastz end-of-file

Housekeeping

Shortcuts for Yasra

There are several shortcut options to support the Yasra mapping assembler. These provide canned sets of option settings that work well for aligning an assembled reference sequence (as the target) with a set of shotgun reads (as the query). They are selected based on the expected level of identity between the sequences. For example, ‑‑yasra90 should be used when we expect 90% identity. The ‑‑yasraXXshort options are appropriate when the reads are very short (less than 50 bp).

Occasionally, newer releases of LASTZ change the Yasra shortcut options. This is done as an improvement, so most users will want to use the shortcuts shown above. Hoever, in order to support backward compatibility for users that want to reproduce previous results, all previous versions of the shortcuts are included. The syntax is ‑‑<shortcut>:<version>, where <version> is the LASTZ version number that contained the shortcut.

Help

Sequence Specifiers

A target or query sequence specifier normally just indicates a file to be used in the alignment; however various pre-processing actions can also be specified. These are performed as the sequences are read from the file, and may include selecting a particular sequence and/or subrange, masking, adjusting sequence names, etc.

The format of a sequence specifier is

    <file_name>[[<actions>]]*

The <file_name> field is required; the actions list is optional. Note that the <actions> are enclosed in literal square brackets (in addition to the meta ones that just indicate they are optional), and consist of a comma-separated list (with no spaces), e.g. [action1,action2,...]. The * indicates that several action lists can be appended; they are treated the same as if they were in a single list.

Alternatively, actions can be specified with the commands ‑‑action:target=<action> and ‑‑action:query=<action>. This allows actions to be set without using square brackets (square brackets are problematic in some command shells).

Note that the actions apply to every sequence in the file. For example, if you specify a subrange of, say, [100..], you will skip the first 99 bp in every sequence.

The following actions are supported:

In addition to the sequence specifier syntax shown above, LASTZ supports a more complicated syntax. This is to maintain compatibility with BLASTZ and early versions of LASTZ. All of the functionality described here can be performed using the newer syntax above.

The complete format of a sequence specifier is

    [<nickname>::]<file_name>[/<select_name>][{<mask_file>}][[<actions>]][-]

As with the simpler syntax, the <file_name> field is required; all other fields are optional. The <file_name> and <actions> fields have the same meaning as in the simpler syntax.

<nickname>:: is equivalent to the <name> field in the nickname=<name> action.

/<select_name> is only valid for the 2Bit file format, and only when the file name ends with ".2bit". It specifies a single sequence from the file to use, rather than all sequences. This is similar to the subset=<names_file> action, except that here a single sequence name is given instead of a file of names. Note that the name must match the mangled sequence name extracted from the file.

{<mask_file>} is identical to the xmask=<mask_file> action.

A - (minus sign) is equivalent to swapping the endpoints in the <subrange> action; it causes the reverse complement of the sequence to be used instead of the sequence itself. Again, this should be used with care, as it can lead to murky interactions with other features. In BLASTZ it was needed for searching only the minus strand, but LASTZ provides a ‑‑strand option for that.

Processing Stages in Detail

Target Sequence Input

The target sequence is read at the beginning and kept in memory throughout the run of the program. Actions such as masking, unmasking, or reverse complement are applied when the file is read. If there are multiple sequences in the target file, they are treated internally as one long sequence (you must use the multiple action in the target file’s sequence specifier to enable this).

In contrast, queries are processed individually and sequentially. Each query sequence is read just before its seeding stage. The seeding through output stages are performed, comparing the query to the target. Then by default, the same stages are repeated to compare the reverse complement of the query to the target, before moving on to the next sequence in the query file.

Scoring Inference

Scoring inference is not normally performed. As described in Inferring Score Sets, LASTZ can iteratively perform the complete alignment process on the target and query, to derive a suitable scoring set. This is only available for special builds of LASTZ, and will usually be too time-consuming to perform for all sequences being aligned. The typical application is to use it once on some sample sequences from the species of interest, save the scoring file, then use that scoring file for subsequent alignments.

Indexing Target Seed Words

This pre-processing stage parses the target sequence(s) into overlapping seed words of some constant length (you can think of these as 12-mers; the actual length is determined by the seed pattern). Each word is converted to a number, called the packed seed word, according to the specified seed pattern. These (word, position) pairs are collected into the target seed word position table. Conceptually, this table is a mapping from a packed seed word to a list of the target sequence positions where that seed word occurs.

This table is one of the major space requirements of the program. Both the memory and time required for seeding can be decreased by using sparse spacing. The ‑‑step option sets a step size: instead of examining every position, seed words are stored only for multiples of the step size. Large step sizes (say, ‑‑step=100) incur a loss of sensitivity, at least at the seeding stage. However, to discover any gapped alignment block we only need to discover one seed (of many) in that alignment, so the actual sensitivity loss is small in most cases. Section 6.2 of [Harris 2007] discusses some experimental results on the effect of step size on the end result.

The presence of biological repeats in the target and query can also be addressed during the building of the position table. A large number of repeats can adversely affect the speed of the program, by increasing the number of irrelevant alignments the program considers in the early stages. LASTZ has three techniques for dealing with repeats.

1. Bases in the target and/or query sequences can be marked as repeats in advance, by using lower case. Target and query words containing lower case bases are left out of the seed word position table and skipped during seeding, respectively, so they do not participate in the seeding stage.
2. If repeat locations are not known, the option ‑‑maxwordcount can be used to remove frequently occurring target seed words from the position table before query processing begins.
3. Dynamic masking (‑‑masking) can be used to mask target positions that have occurred in too many alignments; however this only affects subsequent query sequences.

Seeding

Seeds are short near-matches between target and query sequences. They identify likely regions of homology that warrant further investigation, and serve as starting points for bootstrapping the alignment process. "Short" typically means less than 20 bp. Early alignment programs used exact matches (e.g. of length 12) as seeds, but more recent programs have used spaced seeds (these are described in more detail in the Seed Patterns section). For the purposes of this section, a seed can be thought of as a 12-mer exact match.

To locate seeds, the query sequence is parsed into seed words the same way the target is (except that ‑‑step does not apply to the query; we look at every seed word). Each packed seed word is used as an index into the target seed word position table to find the target positions that have a seed match for this query position. Query seed words containing lower case bases are skipped, so that repeats will not participate in the seeding stage.

Quantum Seeding:

The quantum seeding threshold can also be set as a percentage of the maximum word score possible. If an exact match seed is used, the maximum word score is the highest value in the substitution matrix multiplied by the seed length. If a spaced seed is used, the multiplier is the number of 1 positions in the pattern.

Note that the seeding options that provide special treatment for transitions (Ts in the seed pattern, half-weight seeds, allowing one or two match positions to be transitions, etc.) are not supported for quantum alignments. These options would make the quantum seeding procedure more complex, and are not really necessary because the quantum mechanism itself provides an alternative way to increase the alignment sensitivity. Also note that q-words containing lower case bases are not discarded, since the quantum alphabet is arbitrary and many ASCII bytes do not even have upper/lowercase versions.

Gap-free Extension

In this stage, each seed is extended without allowing gaps to determine whether it is part of a high-scoring segment pair (HSP). The seed is extended along its DP matrix diagonal independently in both directions according to an extension rule, currently either exact match, M-mismatch, or x-drop.

Exact match extension (‑‑exact) simply extends the seed until a mismatch is found. If the resulting length is enough, the extended seed is kept as an HSP for further processing. Exact match extension is most useful when the target and query are expected to be very similar, e.g. when aligning short reads to a similar reference genome.

M-mismatch extension (‑‑<M>mismatch) extends the seed to find the longest interval that includes the entire seed and contains no more than M mismatches. If the resulting length is enough, the extended seed is kept as an HSP for further processing. M-mismatch extension is most useful when the approximate divergence between the target and query is known, and HSPs of a known length are desired. It provides a way to specify both length and identity thresholds together, with more flexibility than ‑‑exact.

In x-drop extension (‑‑xdrop), as we extend in each direction we track the cumulative score for the extended match according to the substitution scoring matrix. The extension is stopped when the score drops off by more than the given x-drop threshold; that is, when the difference between the peak score seen so far and the current score is more than <dropoff>. (Another way to think of it is that the segment ends when a section scoring worse than −<dropoff> is encountered.) The extension is then trimmed back to the peak point. If the combined score of the seed plus both extensions meets the threshold set by the ‑‑hspthresh option, it qualifies as an HSP and is kept for further processing. Matches that do not meet the score threshold are discarded. The ‑‑entropy options control whether or not the scores are adjusted for nucleotide entropy when they are compared to the threshold.

Adaptive Score Threshold:

Diagonal Hashing:

HSP Chaining

The chaining stage aims to find a series of HSPs that forms a high-scoring path through the DP matrix, aligning as much as possible while avoiding backtracking in either sequence. Conceptually it does this by examining all combinations of HSPs and scoring the chains according to the relative positions of the HSPs (e.g. the distances between them along the diagonal and anti-diagonal) as well as their individual scores. All HSPs not in the highest-scoring chain are discarded.

Ideally this process selects the "real" alignments, filtering out noise (such as extra alignments due to repeats), and producing a set of HSPs where each base is aligned at most once; however this is not guaranteed. LASTZ’s implementation is primarily intended for the case where elements are known to appear in the same relative order and orientation in the query as in the target. (However, note that because the forward and reverse strands are processed in separate pipelines, it will not necessarily cause inversions to be discarded.) If LASTZ’s implementation of chaining is not suitable, it is possible to substitute another chaining program by first running LASTZ with the ‑‑nogapped and ‑‑writesegments options to get the HSPs, running a separate chaining program to filter them, and then running the final stages of LASTZ on that output via the ‑‑segments option.

Figure 5(a) shows an alignment without chaining, while 5(b) shows the same alignment with chaining.

lastz target query --nochain

lastz target query --chain

Gapped Extension

Before the HSPs are extended further by allowing gaps, each HSP is first reduced to a single anchor point; this allows for the possibility that the optimal alignment may include gaps within the region occupied by the HSP. The gap-free HSP is only an indication of likely homology in that vicinity; other paths through the same region that allow gaps may have a higher score, so we don't want to just extend from the ends of the HSP. Instead we run the gapped algorithm from a single point that we think is most likely to lie on the optimal path, namely the middle of the highest-scoring 31-bp interval in the HSP. A more general (and expensive) approach would be to examine all paths through the square region defined by the HSP, instead of starting from a single anchor point.

Figure 6(a) illustrates the relationship of seeds, HSPs, and anchors. Heavy lines are seeds, which were extended without gaps (see Overview) to create HSPs (thin lines). Blue dots are anchors. Seeds with no HSP shown (gray lines) had low-scoring extensions and were discarded at the gap-free extension stage.

The anchors are then processed in the order of their HSP’s score (highest first). Gapped extension is performed independently in both directions from the anchor point, and the two resulting alignments are joined at the anchor. If the total score meets the threshold specified by the ‑‑gappedthresh option, the joined alignment is kept and passed to the next stage; otherwise it is discarded. If the extension from one anchor happens to go through one or more other anchors, the redundant anchors are dropped from the list.

Figure 6(b) shows the relationship of anchors and their gapped extensions. The blue dots are the anchors from 6(a), which are extended in both directions to form gapped alignments (squiggly lines; the gaps are too small to be visible at this scale). One anchor had low-scoring extensions that did not meet the threshold. Another had an extension that ran directly through a nearby anchor; that anchor did not need to be processed separately.

The gapped extensions are computed using a typical dynamic programming recurrence for affine gap alignment (e.g. [Myers 1989] or [Gusfield 1997]), beginning at the anchor and terminating at the point with the highest cumulative score. The portion of the DP matrix examined is reduced by disallowing low-scoring regions (see [Zhang 1998]): wherever the alignment score drops below the peak score seen so far by more than the threshold specified in the ‑‑ydrop option, the DP matrix is truncated and no further cells are computed along that row or column. By default the extension is then trimmed back to the location of the peak score; thus the extension normally ends when all remaining sub-alignment possibilities (paths in the DP matrix) begin with sections that score worse than −<dropoff>. However for alignments where the extension reaches the end of the sequence, you can suppress this trimming by specifying the ‑‑noytrim option, which is recommended when aligning short reads.

Figure 7 shows the effect of the y-drop threshold in more detail. Extension is performed in two directions from the anchor (in this example, to the upper right and lower left, because both sequences are on the positive strand). The gray region is the portion of the DP matrix explored by the extension algorithm; its boundary is formed by the points where the score dropped from the maximum by more than the y-drop threshold.

Back-end Filtering

Whatever alignment blocks have made it through the above gauntlet are then subjected to identity, continuity, coverage and match count filtering (as specified by the ‑‑filter=identity, ‑‑filter=continuity, ‑‑filter=coverage, ‑‑filter=nmatch, ‑‑filter=nmismatch, ‑‑filter=ngapand ‑‑filter=cgap options, respectively). Blocks that do not meet the specified range for each feature are discarded.

Identity is the fraction of aligned bases (excluding columns containing gaps or non-ACGT characters) that are matches, expressed as a percentage. The numerator is the number of matches in the alignment block, while the denominator is the number of matches plus the number of mismatches. Characters that differ only in upper vs. lower case are counted as matches. Columns containing gaps or non-ACGT characters play no part in this computation, and it is independent of the settings for ‑‑ambiguous=n and bad_score. Identity cannot be determined for alignments with quantum DNA, because of the potential ambiguity of the symbols.

Continuity is the fraction of alignment columns that do not contain gaps, expressed as a percentage. The numerator is the number of matches plus mismatches in the alignment block, while the denominator is the number of columns. Unlike the computation of identity, here "matches plus mismatches" includes all non-gap columns regardless of whether they contain non-ACGT characters.

Coverage is the fraction of bases in the entire input sequence (target or query, whichever is shorter) that are included in the alignment block, expressed as a percentage. Such bases are aligned in the block to either bases or gaps in the other sequence. Note that if there are multiple sequences in the target and/or query, only the current one is considered; however if an input sequence is spliced with runs of Ns or Xs, then the combination of all its subsequences (including the splice characters between them) is considered as one input sequence, because LASTZ does not explicitly recognize the splicing. Further, if a separator character is used, again the combination of all subsequences is considered as one input sequence (including the separator characters). Also note that each block’s coverage is computed independently of other blocks, and each must meet any specified filter range by itself; blocks cannot be combined to meet coverage requirements.

Match Count, or nmatch, is the number of matched bases in the alignment. Characters that differ only in upper vs. lower case are counted as matches, columns containing gaps or non-ACGT characters are not. Match count cannot be determined for alignments with quantum DNA, because of the potential ambiguity of the symbols.

Mismatch Count, or nmismatch, is the number of aligned bases in the alignment that are not matches. This includes substitutions as well as non-ACGT characters (even if they are identical), but not gaps. Mismatch count cannot be determined for alignments with quantum DNA, because of the potential ambiguity of the symbols.

Gap Count, or ngap, is the number of gaps in the block, counting each run of gapped columns as a single gap.

Gap Column Count, or cgap, is the number of gaps in the block, counting each gapped column as a separate gap.

Interpolation

Once the above stages have been performed, it is not uncommon to have regions left over in which no alignment has been found. In the interpolation stage (activated by the ‑‑inner option) we repeat the seeding through gapped extension stages in these leftover regions, at a presumably higher sensitivity. Using such high sensitivity from the outset would be computationally prohibitive (due to the excessive number of false, low-scoring matches), but is feasible on the smaller, leftover regions.

Another complete alignment round (seeding, gap-free extension, chaining, and gapped extension, even if some of these were skipped in the main alignment; but not back-end filtering) is performed in the small areas between the alignment blocks found in the preceding main alignment stage. Only regions within 20K bp from the endpoints of the passed-in alignment blocks are searched. Seeding for this alignment requires a 7-bp exact match with no transitions, and uses the specified scoring threshold for both its gap-free and gapped extension sub-stages. (This threshold should generally be set lower than the corresponding ones in the main alignment, in order to increase the sensitivity of the interpolation.) All other parameters are the same as those used for the main alignment stages.

Figure 8 shows the operation in more detail. The alignment blocks resulting from gapped extension are shown in 8(a) as squiggly lines. After interpolation, in 8(b), additional alignment blocks have been discovered in the red areas. Note that there are still some holes remaining, where these sequences just don't align well.

lastz target query

lastz target query --inner=1000

Alignment Output

The alignment blocks found by the preceding pipeline of stages are written to stdout (or to a file specified with the ‑‑output option) in the requested format. These may be seeds, gap-free HSPs, or gapped local alignments, depending on which stages were performed. There is no particular order to the alignment blocks for an individual query sequence (e.g. they are not sorted by score or position). However, since the query sequences are processed serially, the blocks for each one will appear together in the output.

File Formats

LASTZ typically receives two sequence files and possibly a scoring file as inputs, and produces an alignment file as output.

DNA sequences can be provided in FASTA, FASTQ, Nib, or 2Bit format, or indirectly via an HSX index. These sequences contain a series of A, C, G, T, and N characters in upper or lower case. Lower case indicates repeat-masked bases, while Ns represent unknown bases if the ‑‑ambiguous=n option is specified. (By default, a run of Ns or Xs is used to separate sequences that have been catenated together for processing, but this is now deprecated; see Non-ACGT Characters, Splicing, and Separation for a discussion of the use of Ns and Xs.) As an alternative to DNA sequence, quantum DNA using an abstract alphabet can be used as the query (but not as the target).

The FASTA, FASTQ, 2Bit and HSX formats support more than one sequence within the same file. Files containing multiple sequences are normally only used as the query; however invoking the multiple action in the file’s sequence specifier allows them to be used for the target as well. Also, the subset action allows one or more sequences to be selected from such a file.

The FASTQ format carries base-calling quality values as well as DNA.

FASTA (sequence input)

FASTA format stores DNA sequences as plain text. The first line begins with a > followed by the name of the sequence, and all subsequent lines contain nucleotide characters. The lines can be of any length. If the file contains multiple sequences, each should start with its own > header line. NCBI FASTA specification

Note that although the official FASTA specification allows the character X only in amino acid sequences, LASTZ accepts it in DNA sequences as a splicing character. However, LASTZ does not currently support IUPAC-IUB ambiguity codes other than N (such as R, W, etc.), beyond the treatment afforded by ‑‑ambiguous=iupac.

A special case, non-conforming to the official standard, is made to allow a special user-specified separator character. Usually this will be N or X, but any other printable ASCII character that suits the user’s needs is acceptible.

It has become common for suppliers of FASTA files to pack a plethora of additional information into a sequence’s header line. This extra information can create difficulties for many sequence processing tools. For example, headers often contain spaces but file formats such as MAF do not allow spaces in sequence names. To compensate for this, LASTZ provides several options for extracting a concise name from sequence headers; see Sequence Name Mangling for details.

FASTQ (sequence input)

FASTQ format stores DNA and base-calling quality sequences as plain text, and is primarily used to describe the results of short-read sequencing runs. As explained in [Cock 2009], this format has evolved over time in the Bioformatics community. LASTZ only supports a subset of this format, prohibiting line-wrapping within DNA or quality sequences.

Each sequence consists of four lines. The first line begins with a ‑ followed by the name of the sequence. The second line contains nucleotide characters. The third line begins with a +, optionally followed by the name of the sequence (which, if present must match that of the first line). The fourth line contains quality characters.

There are several conflicting standards for encoding quality values in FASTQ files, but (as of this writing) the differences are not relevant to LASTZ. LASTZ currently does not make any computational use of the qualities, and simply copies them into the output file when appropriate.

LASTZ treats IUPAC-IUB ambiguity codes in FASTQ files the same as those in FASTA files.

Nib (sequence input)

Nib format stores a single unnamed DNA sequence, packed as two bases per byte. UCSC Nib specification

2Bit (sequence input)

2Bit format stores multiple DNA sequences, encoded as four bases per byte with some additional information describing runs of masked bases or Ns. UCSC 2Bit specification

Sequence names in 2Bit files have all the same problems as in FASTA files, so Sequence Name Mangling applies to these files as well.

Quantum DNA (sequence input)

A quantum DNA file describes a single sequence of "quantum" DNA, which uses an abstract, user-defined alphabet. Each position in the sequence is a byte with a value in the range 0x01..0xFF, which can represent an ambiguity code, amino acid, or any other meaning you desire. LASTZ does not try to interpret these in any way; it just aligns them as abstract symbols corresponding to columns in the scoring matrix. Note that the value 0x00 is prohibited.

The file itself is stored in a binary format described by the table below. It can be written on either a big-endian or little-endian machine; LASTZ determines the byte order of multi-byte fields by examining the magic number at the start of the file. Be sure to use the quantum action in the file’s sequence specifier to notify LASTZ that it contains quantum DNA.

Quantum Code File

This file is used with the quantum action in a sequence specifier. It defines a mapping from quantum DNA symbols to vectors of values for the four nucleotides A, C, G, and T. Usually these indicate the nucleotide probability distribution for each symbol in the quantum alphabet. However, LASTZ doesn't interpret the values, and only uses them to to augment the display of alignment blocks in the Human-Readable Text output format.

Each line in the file gives the mapping for one symbol. Lines beginning with a # are considered to be comments and are ignored, as are blank lines. Data lines have five columns, separated by whitespace. The first field contains the symbol, as either a single character or two hexadecimal digits, while the remaining four fields contain values for A, C, G, and T, respectively. Each value can be either a single floating-point number or a fraction (two floating-point numbers with a / between them, without spaces). Any symbols in the quantum alphabet that aren't listed in this file receive zeroes for all four values.

Here is an example.

    # sym p(A|sym) p(C|sym) p(G|sym) p(T|sym)
      01  0.125041 0.080147 0.100723 0.694088
      02  0.111162 0.053299 0.025790 0.809749
      03  0.065313 0.007030 0.004978 0.922679
       ... more rows here ...
      FF  0.209476 0.014365 0.755682 0.020477

Sequence Name File

This file is used with the subset action in a sequence specifier to select particular sequences for processing. It consists of one sequence name per line. Lines beginning with a # are considered to be comments and are ignored, as are blank lines. Only the first whitespace-delimited word in any line is read as the name; the rest of the line is ignored.

Note that when used in conjunction with a FASTA or 2Bit file, the names must appear in the same order as they appear in the corresponding sequence file, and must match the mangled name extracted from that file. When used with an HSX file, the names can be in any order but must match names indexed in the HSX file.

Sequence Masking File

This file is used with the xmask and nmask actions in a sequence specifier. It can also be created by using the ‑‑outputmasking=<file> or ‑‑outputmasking:soft=<file> options. It consists of one interval per line, without sequence names. Lines beginning with a # are considered to be comments and are ignored, as are blank lines. Only the first two whitespace-delimited words in any line are interpreted as the interval; the rest of the line is ignored.

Each interval describes a region to be masked, and consists of

    <start> <end>

Here is an example. If the target sequence is hg18.chr1, this would mask the 5' UTRs from several genes. Note that the third column is neither required nor interpreted by LASTZ, and acts as a comment.

     884484  884542  NM_015658
     885830  885936  NM_198317
     891740  891774  NM_032129
     925217  925333  NM_021170
     938742  938816  NM_005101
     945366  945415  NM_198576
    1016787 1016808  NM_001114103
    1017234 1017346  NM_001114103
    1041303 1041486  NM_001114103

Sequence Masking File, Three Fields

This file is created by using the ‑‑outputmasking+=<file> or ‑‑outputmasking+:soft=<file> options. It consists of one interval per line, with sequence names.

Each interval describes a region that has been masked, and consists of

    <name> <start> <end>

Scoring File

This file is used with the ‑‑scores option to specify a set of (mostly) scoring-related parameters en masse. The score set consists of a substitution matrix and other settings. The other settings come first and are individually explained in the table below. All settings are optional, and most of them have exact correspondence to command-line options and the same defaults (unless otherwise specified in the table). Command-line settings always override settings in this file. Any line may end with a comment (# is the comment character).

In the matrix, rows correspond to characters in the target sequence, while columns correspond to characters in the query. Matrix labels can be specified either as single ASCII characters or as two-digit hexadecimal values in the range 01..FF (do not add a leading 0x). Note that the value 00 is not allowed. The rows and columns of the matrix need not have the same set of labels, so for example, a matrix might describe scoring between the 4-letter DNA alphabet and the 15-letter ambiguity alphabet. Any labels other than A, C, G, and T (or their hex equivalents) are treated as quantum DNA.

Score values can be floating-point if the lastz_D version of the executable is used instead of lastz.

Here is an example:

    # This matches the default scoring set for BLASTZ
    
    bad_score          = X:-1000  # used for sub['X'][*] and sub[*]['X']
    fill_score         = -100     # used when sub[*][*] is not defined
    gap_open_penalty   =  400
    gap_extend_penalty =   30

         A     C     G     T
    A   91  -114   -31  -123
    C -114   100  -125   -31
    G  -31  -125   100  -114
    T -123   -31  -114    91

BLASTZ scoring files are also accepted. These only contain a substitution matrix, and row labels must be absent (they are assumed to be the same as the column labels). No other settings are allowed.

       A     C     G     T
      91  -114   -31  -123
    -114   100  -125   -31
     -31  -125   100  -114
    -123   -31  -114    91

Inference Control File

When LASTZ is asked to infer substitution scores and/or gap penalties from the input sequences (e.g. via the ‑‑infer option), this file is used to set parameters that control the inference process.

Here is an example (note that currently the parsing of this file is less flexible than some of the others, and * is the only arithmetic operator supported):

    # base the inference on alignments in the middle half by identity
    min_identity       = 25.0%    # 25th percentile
    max_identity       = 75.0%    # 75th percentile

    # scale scores so max substitution score will be 100, and only use
    # alignments scoring at least as well as 20 ideal matches
    inference_scale    = 100      # max substitution score
    hsp_threshold      = 20*inference_scale
    gapped_threshold   = hsp_threshold

    # allow substitution score inference to iterate at most 20 times;
    # don't perform gap penalty inference -- instead hardwire gap penalties
    # relative to max substitution
    max_sub_iterations = 20
    max_gap_iterations = 0
    gap_open_penalty   = 4*inference_scale
    gap_extend_penalty = 0.3*inference_scale

    # use all seedword positions (don't sample)
    step               = 1

    # adjust for entropy when qualifying HSPs
    entropy            = on

min_identity and max_identity specify the range of sequence identity upon which inference is based; only alignment blocks within this range contribute to the inference. If the value ends with a percent sign, it represents a percentile of the identity distribution over all the blocks; otherwise it is a fixed percent identity value. For example, min_identity=70 and max_identity=90 indicates that blocks with identity ranging from 70 to 90 percent will be used, while min_identity=25% and max_identity=75% indicates that half of the blocks will be used (from the middle of the distribution). The defaults are min_identity=0 and max_identity=100 (i.e., no blocks are excluded from inference due to percent identity).

inference_scale specifies a value for the largest substitution score (i.e., the score for the best match). All other scores are scaled proportionally. If this is set to none, the scores will be log-odds using base 2 logarithms. The default is inference_scale=100.

hsp_threshold and gapped_threshold correspond to the command line ‑‑hspthresh and ‑‑gappedthresh options. The defaults are hsp_threshold=3000 and gapped_threshold=hsp_threshold.

max_sub_iterations and max_gap_iterations specify limits on the number of inference iterations that will be performed. For example, if you only want a substitution scoring matrix, you can set max_gap_iterations=0. The defaults are max_sub_iterations=30 and max_gap_iterations=0.

gap_open_penalty and gap_extend_penalty correspond to the command line ‑‑gap=[<open>,]<extend> option. These are used for the first iteration of gap-scoring inference. The defaults are gap_open_penalty=3.25*worst_substitution and gap_extend_penalty=0.24375*worst_substitution.

step corresponds to the command line ‑‑step option. A large step, e.g. step=100, could potentially speed up the inference process. Ideally, this would base the inference on a sample of only one percent of the whole. However, the sample actually ends up larger than that and is biased toward HSPs that are either longer or have a lower substitution rate. This happens because sampling occurs at the seed level, and such HSPs generally have more seeds. Future versions of LASTZ may include a means to compensate for this bias. The default is step=1.

entropy corresponds to the command line ‑‑entropy option. Legal values are on or off. If on, sequence entropy is incorporated when filtering HSPs. The default is entropy=on.

The value of worst_substitution cannot be set directly. Instead, it is computed from the initial scoring matrix. It is the minimum score in the scoring matrix for any of the symbols A, C, G or T (equivalently, the most negative score or the maximum penalty).

Note that these parameters apply to the inference process only. If the corresponding command line options are also set, those will apply for the final, "real" alignment stages (and will also override the inferred values if there is a conflict), but will not affect the inference itself. Inference cannot be used in conjunction with a scores file.

HSX (Hashed Sequence Index)

An HSX file is an index of sequences in other files, allowing fast random access to those sequences. The current implementation of LASTZ only supports indexing FASTA files. Future versions may include Nib and 2Bit sequences. The following is a brief overview of the file format. For more detailed information, see the HSX specification

An HSX file can be created using the build_fasta_hsx.py utility (included in the tools directory of the LASTZ distribution), using a command like this:

    build_fasta_hsx sequences.fa [more_sequences.fa ...] > index.hsx

It is important that the HSX file has the extension .hsx and resides in the same directory as the files being indexed. Further, the files being indexed must have the extension .fa or .fasta. These rules allow LASTZ to determine the sequence file type when it reads the HSX file, and to locate the files containing the sequences.

The index file includes names to be used for the sequences, which do not have to match the original names or headers in the sequence files. This feature obviates the need for LASTZ to perform sequence name mangling, so most of those actions are not supported for HSX files. Instead, it is the responsibility of the program that creates the index to select suitable names.

Target Capsule File

A target capsule file is essentially a memory dump of several internal data structures related to the target sequence and the target seed word position table. At the present time the authors do not wish to create an official specification for this format, but please see Using Target Capsule Files for information on how to create and utilize them.

Alignment Chores File

A chores file describes a list of sequence interval pairs, indicating that the alignment process is to be restricted to those intervals.

The file contains two intervals per line, one from the target and one from the query, with sequence names. Optionally, the query strand can be specified, as well as an identifying tag. Lines beginning with a # are considered to be comments and are ignored, as are blank lines. # can also be used to put comments at the end of lines, but must be preceeding by whitespace.

Each line looks like

    <name1> <start1> <end1> <name2> [<start2> <end2>] [<strand2>] [id=<tag>] [#<comment>]

When the target name is irrelevant (i.e. when there is only one name in the target sequence file), * can replace <name1>. Similarly, if we don't have a target (or query) subrange, * * can be used in place of start and end. Note that the query subrange and strand are optional, as is the tag. When the strand is not specified, both strands are searched.

Locations are one-based and inclusive on both ends, i.e. origin-one, closed (thus the interval "154 228" has length 75 and is preceded by 153 bases in its sequence). All target intervals are on the positive strand. All query intervals are counted along the forward strand, regardless of which strand is specified.

Target sequence names may appear in any order. Sequence names for the query must appear in the same order as they do in the query file. Because alignment output ordering is on a chore-by-chore basis, it is good practice to include all positive strand intervals for a query before any negative strand intervals for that query. Some downstream tools may depend on this ordering.

The tag can be any short string the user wants to associate with the chore (excluding whitespace). As of this writing, the only use of the tag field is that it can be copied to the output file by use of the chore field for ‑‑format=general.

Here is an example.

    chr9  116517410 116518409  READ_00070 *   *   + id=DFZ
    chr3  157707345 157708344  READ_00070 *   *   + id=EDZ
    chr9  112944437 112945436  READ_00078 101 200 + id=FAC
    chr1  3377578   3378577    READ_00078 *   *   + id=LLH
    chr2  175604671 175605670  READ_00078 *   *   - id=DFZ
    chr2  230613705 230614704  READ_00079           id=DFZ
    chr9  20387422  20388421   READ_00355 *   *   + id=DFZ
    chr8  16396215  16397214   READ_00355 *   *   + id=MNQ
    chr14 *         *          READ_00355 *   *   - id=MNQ
    chr4  50534096  50535095   READ_00355 *   *   - id=QOY
    chr6  58308766  58309765   READ_00376 *   *   - id=EDZ
    chr5  172249269 172250268  READ_00376 *   *   - id=FAC
    chr9  123860065 123861064  READ_00376 *   *   - id=MNQ

Segment File

A segment file describes a list of segments representing gap-free alignments. This list is either produced internally by LASTZ as a result of the gap-free extension stage (see Overview), or read from a user-supplied file via the ‑‑segments option. The latter causes LASTZ to skip the indexing, seeding, and gap-free extension stages and begin with the chaining stage (or the next specified stage, if chaining is not requested).

The file contains two intervals per line, one from the target and one from the query, with sequence names. Lines beginning with a # are considered to be comments and are ignored, as are blank lines. # can also be used to put comments at the end of lines.

Each line looks like

    <name1> <start1> <end1> <name2> <start2> <end2> <strand2> [<score>] [#<comment>]

Locations are one-based and inclusive on both ends, i.e. origin-one, closed (thus the interval "154 228" has length 75 and is preceded by 153 bases in its sequence). Negative strand intervals are measured from the 5' end of the query’s negative strand (corresponding to the rightmost end of the given query sequence, i.e. counted along the reverse strand). All target intervals are on the positive strand. The two intervals must have the same length (since these alignments are gap-free). The score is used to determine the processing order during gapped extension. Segments without scores are given a score of zero.

Query sequence names must appear in the same order as they do in the query file. For each query sequence, normally all positive strand intervals must appear before any negative strand intervals. Sequence names for the target may appear in any order, and are only meaningful if the multiple action is used; otherwise they are ignored. Intervals with names not found in the target or query are not allowed. In cases where sequence names are either unknown or of no importance (e.g. when all sequences in the file have the same name), a * can be used as a generic sequence name.

Here is an example.

    R36QBXA37A3EQH 151 225  Q81JBBY19D81JM 14  88 +  6875
    R36QBXA37D4L6V  26 100  Q81JBBY19D81JM 10  84 +  6808
    R36QBXA37EVLNU  19  93  Q81JBBY19D81JM  7  81 +  6842
    R36QBXA37CEBPD   8  81  Q81JBBY19D81JM  9  82 +  7108
    R36QBXA37BLO6X 132 205  Q81JBBY19D81JM 11  84 -  7339
    R36QBXA37A2W3P 162 214  Q81JBBY19D81JM  2  54 -  5024
    R36QBXA37A9395  62 136  Q81JBBY19A323K 18  92 +  7231
    R36QBXA37DNC74  18  82  Q81JBBY19A323K  2  66 +  6418
    R36QBXA37CTR26  83 167  Q81JBBY19ASA7F 19 103 +  8034
    R36QBXA37C2TAC  95 181  Q81JBBY19ASA7F 15 101 +  8272

LAV (alignment output)

LAV is the format produced by BLASTZ, and is the default. It reports the alignment blocks grouped by "contig" (chromosome, scaffold, read, etc.) and strand, and describes them by listing the coordinates of gap-free segments. This format is compact because it does not include the nucleotides, but consequently interpretation usually requires access to the original sequence files, and it is not easy for humans to read. LAV specification (same specification at PSU)

The option ‑‑format=lav+text adds textual output for each alignment block (in the same format as the ‑‑format=text option), intermixed with the LAV format. Such files are unlikely to be recognized by any LAV-reading program.

AXT (alignment output)

AXT is a pairwise alignment format popular at UCSC and PSU. UCSC AXT specification

The option ‑‑format=axt+ reports additional statistics with each block, in the form of comments. The exact content of these comment lines may change in future releases of LASTZ.

MAF (alignment output)

MAF is a multiple alignment format developed at UCSC. The MAF files produced by LASTZ have exactly two sequences per block: the first row always comes from the target sequence, and the second from the query. UCSC MAF specification

The option ‑‑format=maf+ reports additional statistics with each block, in the form of comments. The exact content of these comment lines may change in future releases of LASTZ.

The option ‑‑format=maf- suppresses the MAF header and any comments. This makes it suitable for concatenating output from multiple runs.

UCSC’s MAF should not be confused with other formats that have the same name. For example, the MIRA sequence assembler project has a file format named MAF, but it is a completely unrelated file format and is not supported by LASTZ.

SAM (alignment output)

SAM is a pairwise alignment format used primarily for short-read mapping, and supported by the SAMtools programming suite. This format is described in [Li 2009], and as of May 2011 a specification for it can be found at the SAMtools page at SourceForge.

For SAM files, LASTZ assumes that the target sequence is the reference and that query sequence(s) are short reads. For alignments that don't reach the end of a query, ‑‑format=sam uses "hard clipping", while ‑‑format=softsam uses "soft clipping". See the section on "clipped alignment" in the SAM specification for an explanation of what this means.

The options ‑‑format=sam- and ‑‑format=softsam- suppress the SAM header lines. This makes them suitable for concatenating output from multiple runs.

CIGAR (alignment output)

CIGAR is an acronym for Concise Idiosyncratic Gapped Alignment Report, a pairwise alignment format defined originally by the Exonerate alignment program. The format has since been adapted in different forms, as ensembl cigar format and as an extended cigar string in SAMtools. For ‑‑format=cigar, LASTZ implements Exonerate CIGAR. LASTZ implements other CIGAR variants for ‑‑format=sam and as fields for ‑‑format=general.

Exonerate CIGAR format does not include nucleotides; instead it describes the locations of indels (but not substitutions) using run-length encoding. An alignment is characterized as runs of M (match and/or substitution), I (query contains a base not in target), and D (target contains a base not in query). Each run is encoded by the letter code, whitespace, and the length; multiple runs are separated by whitespace. The format also includes positional information for the start of the alignment. An example is shown at the end of this section. While there seems to be no complete, definitive specification for CIGAR, the CIGAR files produced by LASTZ are believed to match the format produced by Exonerate.

In the other variants of CIGAR, whitespace is removed and the order of the letter code and length are reversed (length appears before letter code). In some variants the length is omitted if it is 1; in other variants M runs are divided into = (match) and X (substitution). SAMtools extended cigar strings allow S and H runs to describe clipping operations for short sequences. LASTZ implements combinations of these variants where appropriate; details are described in ‑‑format=general:cigar, ‑‑format=general:cigarx and ‑‑format=sam.

    target:  ...GATTAAGAGTCTGTCCGACCTTCTTCT---GGGTTTACCGAAGCCCACTTAGCTGATATTCGA...
                   ||||||||||||||||X|||||||   |||||||  X||||||||||||||||||
     query:     ACCTAAGAGTCTGTCCGACATTCTTCTACGGGGTTTA--TAAGCCCACTTAGCTGATAAGGTT
                   ↑      1         2         3           4         5    ↑    6
                0123456789012345678901234567890123456--789012345678901234567890

For ‑‑format=cigar, the alignment would be described by this line:

    cigar: query 3 56 + target <start> <end> <strand> <score> M 24 I 3 M 7 D 2 M 19

For ‑‑format=general:cigar, the alignment path would be described by this field:

    24M3I7M2D19M

For ‑‑format=general:cigarx, the alignment path would be described by this field:

    16=X7=3I7=2DX18=

For ‑‑format=sam, the alignment path would be described by this field:

    3H24M3I7M2D19M5H

BLASTN (alignment output)

The BLASTN format reports pairwise alignments in a format similar to NCBI’s BLASTN program. Output is modeled upon version 2.2.24+ of the standalone version of BLASTN available from NCBI’s BLAST ftp site. Output should be similar that produced by the command

    blastn -task blastn -db <target> -query <query> -outfmt 7

The format is tab-delimited with one alignment reported per line, plus an additional header. Here is some sample output:

    # lastz --format=blastn
    # Query: orange
    # Database: apple
    # Fields: query id, subject id, % identity, alignment length, mismatches, gap opens, q. start, q. end, s. start, s. end, evalue, bit score
    orange apple 82.14  2072 142 67 2    1926 103  2093 0     1972
    orange apple 100.00 14   0   0  1906 1919 2086 2073 0.043 26.5
    orange apple 93.33  15   1   0  1763 1777 2004 1990 0.53  22.9

Most of the fields correspond directly to fields available in the General output format. These are query id=name2, subject id=name1, %identity=blastid%, alignment length=ncolumn, mismatches=nmismatch, gap opens=ngap, q.start=start2, and q.end=end2. The fields s.start and s.end are nearly equivalent to start1 and end1, but when the alignment is to the reverse strand, they appear in the other order (i.e. s.start > s.end).

The two remaining fields, evalue and bit score, are crudely estimated from LASTZ’s score field, but are not strictly correct. Further, these approximations assume that default LASTZ scores are used. Otherwise they are unlikely to be good approximations. The approximation formulas are

     evalue    = 3.0e9*exp(-0.01421*score)
     bit score = 0.0205*score

PAF (alignment output)

The PAF format reports pairwise alignments in a format compliant with minimap2, available from the minimap2 github site. A simple spec for PAF files can be found here, while more detail can be found here.

LASTZ currently doesn't compute mapping quality scores, and reports 255 (missing) for the mapping quality field (column 12).

The PAF spec allows for certain SAM-like tags to be included. LASTZ produces two such tags — AS reports the lastz alignment score and cg reports the alignment's CIGAR string.

Two variations of the format are provided, to support two different formats for the CIGAR field. For ‑‑format=PAF the CIGAR string does not distinguish matches from substitutions, which is what minimap2 produces. For ‑‑format=PAF:wfmash the CIGAR string does distinguish matches from substitutions.

Differences (alignment output)

LASTZ’s Differences format reports each difference between target and query on a separate line, where a difference is any indel or run of mismatches. It is intended for comparisons between close sequences, such as when comparing reads from a human individual to the human reference genome, or reads from a bacterial strain to a reference sequence for the same bacterium. The format is a tab-delimited table with one line per difference; it is well-suited for use with spreadsheets and the R statistical package.

The columns reported in this format are the name, start & end of the difference, strand, and overall size for the target; the name, start & end of the difference, strand, and overall size for the query; the text of the difference in the target, then in the query; and finally the text of the complete alignment block containing the difference, first in the target, then in the query. Intervals are origin-zero, half-open and counted along the forward strand.

The example below compares output in this format to similar results using the General output format for the same input sequences. For the Differences output, column numbers have been added for discussion (they are not in the actual output file). Each line in the output represents slight evidence that a mutation occurred changing the target sequence (chr22 here) to the query sequence (column 6). Columns 11 and 12 indicate the specific mutation that has putatively occurred. For example, the first line suggests that either an A has been inserted into chr22 at position 14485783, or an A has been deleted from EAYGRGI02GQ0SL at position 167 (actually, between positions 166 and 167). Note that there are three differences reported for EAYGRGI02GQ0SL, so it appears on three lines. The fifth line shows a putative SNP at chr22 position 15234401, with a C in the reference and a G in the read, while the seventh line shows evidence for an inversion of neighboring bases (AG vs. GA). Note that there are no lines for EAYGRGI01BIQCW, indicating a perfect match for that block (i.e., no differences).

Sample output for ‑‑format=differences.

     (1)     (2)      (3)  (4)   (5)         (6)       (7) (8) (9) (10) (11)(12)  (13)     (14)
    chr22 14485783 14485784 + 49691432  EAYGRGI02GQ0SL 167 167  +  303   A   -   TGAGA... TGAGA...
    chr22 14485791 14485792 + 49691432  EAYGRGI02GQ0SL 174 174  +  303   A   -   TGAGA... TGAGA...
    chr22 14485843 14485843 + 49691432  EAYGRGI02GQ0SL 225 226  +  303   -   A   TGAGA... TGAGA...
    chr22 14731895 14731895 + 49691432  EAYGRGI01EAV19 228 229  -  298   -   A   CTTCT... CTTCT...
    chr22 15234401 15234402 + 49691432  EAYGRGI02H5ZGS 99  100  -  180   C   G   CGAAT... CGAAT...
    chr22 15255536 15255537 + 49691432  EAYGRGI01BTT7U 56  56   -  267   A   -   TTTGC... TTTGC...
    chr22 15255552 15255554 + 49691432  EAYGRGI01BTT7U 71  73   -  267   AG  GA  TTTGC... TTTGC...
    chr22 15255617 15255618 + 49691432  EAYGRGI01BTT7U 136 136  -  267   A   -   TTTGC... TTTGC...
    chr22 15255624 15255625 + 49691432  EAYGRGI01BTT7U 142 142  -  267   A   -   TTTGC... TTTGC...

Sample output for ‑‑format=general:name1,zstart1,end1,strand1,size1,name2,zstart2+,end2+,strand2,size2,text1,text2.

    chr22 14485616 14485920 + 49691432  EAYGRGI02GQ0SL 0   303  +  303   TGAGA... TGAGA...
    chr22 14731668 14731964 + 49691432  EAYGRGI01EAV19 0   297  -  298   CTTCT... CTTCT...
    chr22 15234302 15234482 + 49691432  EAYGRGI02H5ZGS 0   180  -  180   CGAAT... CGAAT...
    chr22 15238845 15239070 + 49691432  EAYGRGI01BIQCW 0   225  -  225   TGGAA... TGGAA...
    chr22 15255480 15255750 + 49691432  EAYGRGI01BTT7U 0   267  -  267   TTTGC... TTTGC...

(This example aligns reads from the genome of James Watson (available from NCBI’s trace archive by querying for CENTER_NAME = 'CSHL' and CENTER_PROJECT = 'Project Jim') vs. the human reference genome hg18.)

R Dotplot (alignment output)

This is a home-grown format designed to facilitate plotting the alignment blocks with the R statistical package. Alignments are reduced to a series of gap-free segments, each of which is written in three lines as shown below. Endpoints are origin-one, closed, and alignments on the reverse strand have <..._query_end> less than <..._query_start> so that R will draw them in the reverse (slope=−1) orientation.

    <target_name>            <query_name_>
    <segment1_target_start>  <segment1_query_start>
    <segment1_target_end>    <segment1_query_end>
    NA                       NA
    <segment2_target_start>  <segment2_query_start>
    <segment2_target_end>    <segment2_query_end>
    NA                       NA
     ...

The file can then be plotted in R with commands like these:

    dots = read.table("your_file",header=T)
    plot(dots,type="l")

When the the query file contains more than one sequence, alignments for each query sequence are written as shown above. This includes a new header line for each query. Unfortunately the simple R commands shown above will not work to plot a file with more than one query.

When the the target file contains more than one sequence, alignments for target sequences are intermixed in the output file. In this case the entire target is treated as a single sequence, and the target positions reported are relative to this concatenated sequence. This can still be plotted using the simple R commands above, but the target sequences will appear as one concatenated sequence in the plot.

Human-Readable Text (alignment output)

This textual output is intended to be read by people rather than programs. Each alignment block is displayed with gap characters and a row of match/transition characters, and lines are wrapped at a reasonable width to allow printing to paper. The exact format of this output may change in future releases of LASTZ, so programs are better off reading more stable formats like LAV, AXT, or MAF.

General Output (alignment output)

LASTZ’s General format is a tab-delimited table with one line per alignment block and configurable columns. This format is well-suited for use with spreadsheets and the R statistical package, and for filtering with shell commands.

The syntax for this option is:

    ‑‑format=general[:<fields>]

    ‑‑format=general:nmismatch,name1,strand1,start1,end1,name2,strand2,start2,end2

    #nmismatch name1   strand1 start1 end1 name2    strand2 start2 end2
    41         apple8  +       130    930  orange2  -       119    931
    35         apple15 +       113    930  orange3  +       87     909
    52         apple4  +       131    952  orange5  -       111    932
    46         apple7  +       131    930  orange10 +       111    909
    37         apple12 +       131    930  orange11 -       111    909
    38         apple3  +       127    939  orange12 +       107    926

The recognized field names are shown in the table below. Positions (start and end fields) are counted from the 5' end of the aligning strand, unless otherwise indicated in the table. Please see Interval Coordinates for more information about the position numbering systems used in LASTZ.

If the field list is absent, the following fields are printed, in this order:  score, name1, strand1, size1, zstart1, end1, name2, strand2, size2, zstart2, end2, identity, coverage. 

The option ‑‑format=mapping is a shortcut for ‑‑format=general with the following fields:  name1, zstart1, end1, name2, strand2, zstart2+, end2+, identity, coverage, cigarx-.

Field names are normally included as column headers in the first row of the output, preceded by a #. The options ‑‑format=general-[:<fields>] and ‑‑format=mapping- suppress column headers. This makes them suitable for concatenating output from multiple runs.

Other Output

LASTZ includes support for other output formats which are intended mainly for the convenience of the developers. If you have specific questions, please contact us.

Advanced Topics

Aligning to Whole Genomes

Aligning queries to a whole genome can be accomplished in a single run of lastz by using the multiple action in the target file’s sequence specifier. This causes lastz to load all of the target’s sequences into memory. However, sequence indexing inside lastz is limited to 31-bit positions, which limits the overall size of the target to 2 gigabases.

To facilitate larger genomes, an additional executable (lastz_32) can be built. The two executables are basically the same; the only difference is that sequence indexing in lastz is limited to 31-bit positions, while lastz_32 uses 32-bit positions. The use of smaller positions in lastz reduces memory usage and improves performance, but limits the size of the target sequence to 2 gigabases.

To build the lastz_32 executable, enter the following commands from bash or a similar command-line shell (Solaris users should substitute gmake for make). This will build the executable and copy it into your installDir.

    cd <somepath>/lastz-distrib-X.XX.XX/src
    make lastz_32
    make install_32

lastz_32 can then be used as a replacement for lastz in any command line, e.g.

    lastz_32 hg18.fa[multiple] galGal3.fa \
      --notransition --step=20 --nogapped \
      --progress=1 \
      --format=maf > hg18_vs_galGal3.maf

Adjacent Indels

Occasionly the sequences being compared contain unrelated segments of DNA flanked by segments that are related. If the unrelated segments are long enough (and different enough) that two gaps are cheaper than a series of substitutions, the optimal-scoring alignment should contain adjacent indels, like this:

    ...ATAAATTATTATTATTAAATTTTA-------------------CCCCCCCCCCCCCCCCCCTTTTTA...
    ...ATAAATTATTATTATTAAATTTTAGGGGGGGGGGGGGGGGGAG-------------------TTTTA...

However, by default, lastz does not allow an insertion to follow a deletion, or vice versa. So it ends up reporting an alignment like this instead:

    ...ATAAATTATTATTATTAAATTTT------------------ACCCCCCCCCCCCCCCCCCTTTTTA...
    ...ATAAATTATTATTATTAAATTTTAGGGGGGGGGGGGGGGGGA------------------GTTTTA...

The latter alignment doesn't make any sense biologically. However, to maintain backward compatibility with previous versions of LASTZ (and BLASTZ), the default version of LASTZ will produce the latter alignment.

Users that want to allow allow alignments with adjacent indels can build any LASTZ executable with allowBackToBackGaps enabled. This is accomplished by adding allowBackToBackGaps=ON to the make command line, like this:

    make clean
    make lastz_32 allowBackToBackGaps=ON
    make install_32

Interval Coordinates

The biological research community has established several competing standards describing intervals on a strand of DNA. Different programs often use different standards. Since LASTZ supports several input and output formats, it is inevitable that it uses more than one way of describing an interval. We describe the different conventions here.

For this discussion, suppose we have a 50-nucleotide strand of DNA as follows:

        origin-one, closed: 12345678901234567890123456789012345678901234567890
                                      ↓      ↓
                     5' >>> CGACCTTACGATTACCTACTTAACACGTAAACTGAGGGATCAAAAGGAAA >>> 3'
                                      ↑       ↑
    origin-zero, half-open: 01234567890123456789012345678901234567890123456789

Note that since this is DNA it has 5' and 3' ends; we assume that all input sequences follow the standard practice of listing the bases with the 5' end on the left. Here we've highlighted the subsequence ATTACCTA so we can discuss how to describe the interval it occupies. There are two commonly used ways to do this. Both count from 5' to 3' (left to right). One way, origin-one, starts counting from one. The other way, origin-zero, starts counting from zero. So in origin-one, ATTACCTA begins at position 11, while in origin-zero it begins at position 10.

To describe the ending position, there are also two commonly used methods. One way is closed, in which the position of the last nucleotide is given. The other is half-open, in which the position following the last nucleotide is given. These are theoretically independent of the conventions for the origin, but in practice only two of the combinations are commonly used: origin-one, closed and origin-zero, half-open. In the former, ATTACCTA is said to occupy the interval (11,18), while in the latter it is said to occupy the interval (10,18). Notice that only the first number changes between these two paradigms; the second number stays the same.

Another factor to consider is that DNA is usually double stranded, which would look like this:

        along forward:        12345678901234567890123456789012345678901234567890
                                        ↓      ↓
       forward strand: 5' >>> CGACCTTACGATTACCTACTTAACACGTAAACTGAGGGATCAAAAGGAAA >>> 3'
    complement strand: 3' <<< GCTGGAATGCTAATGGATGAATTGTGCATTTGACTCCCTAGTTTTCCTTT <<< 5'
                                        ↑      ↑
        along reverse:        09876543210987654321098765432109876543210987654321

In some cases it makes sense to refer to the interval along the complement strand. For example, if the above sequence was a query and the target contained TAGGTAAT, how should the query position of an alignment of those two be described? One way would be to still refer to the interval along the forward strand (which we also call the plus or positive strand), and just indicate that in fact it was the reverse complement of that interval that aligned. We call this counting along the forward strand. Another way is to count from the other end, from the 5' end of the complement strand (which we also call the reverse, minus or negative strand). We call this counting along the reverse strand, and for clarity we might add "from its 5' end". In this example, if we were using origin-one, closed counting, we would say that TAGGTAAT occurs at (33,40) along the reverse strand. Unless noted otherwise (e.g. for the R Dotplot output format), when counting along the forward or reverse strand LASTZ swaps the interval’s endpoints if necessary, so the position called start is numerically ≤ the position called end. This is a common convention, but there are other programs that leave them unswapped.

Note that when counting positions all characters in the sequence are counted, including runs of Ns or Xs and even invalid characters. This is important so that other programs can use the reported positions to index directly into the original sequences.

Non-ACGT Characters, Splicing, and Separation

The handling of characters other than A, C, G, and T in sequences that are supposed to represent DNA is problematic. In ordinary (non-quantum) DNA sequences, LASTZ currently supports two of these, N and X. They can either be present in the original input file (except that the Nib and 2Bit formats are incapable of containing Xs), or added by using an xmask or nmask action in the sequence specifier. LASTZ can also be configured to tolerate the other IUPAC-IUB ambiguity codes (B, D, H, K, M, R, S, V, W, and Y), and to recognize a special user-specified separator character.

Many database sequences contain Ns to represent bases for which the actual nucleotide is not known (at least, not known with any level of confidence). Ns (or better, Xs) can also be used to mask out regions that have previously been identified as being of no interest, and therefore should not be aligned. And unfortunately, there is also a tradition of using strings of Xs or Ns to splice together multiple sequences to gain efficiency when dealing with programs that were limited to operating on a single sequence.

Although splicing was useful in BLASTZ, it is no longer needed for LASTZ. Since LASTZ can handle multiple target sequences (via the multiple action in the target file’s sequence specifier), it is preferred that users not resort to splicing. If splicing is necessary, the preferred method is to specify a separator character to tell LASTZ explicitely where the splices have occurred.

Replacing BLASTZ with LASTZ in an existing pipeline may still involve spliced sequences, so LASTZ’s default interpretation of non-ACGT characters is the same as BLASTZ’s:  Xs are excluded from the alignment seeding stage, and are so severely penalized by alignment scoring that they will not normally appear in any alignment. Ns are also excluded from seeding, and are penalized about the same as a transversion mismatch. Specifically, any substitution with X is scored as −1000, and any substitution with anything else (other than A, C, G, or T) is scored as −100. Note that you have to put "enough" Xs or Ns between the sequences so that no alignment block will cross the splice. This can be tricky, since gap scoring is only dependent on the length of the gap and not on the characters in the gap. So if a gap the same length as the splice is not penalized more than the y-drop setting, the alignment may hop the splice. As a rough guideline, a splice length of 50 is usually enough with the default settings, but this is not guaranteed.

This default treatment of non-ACGT characters also works well when Xs or Ns are used to mask out regions that should not be aligned. However, it is inappropriate when the sequences contain Ns to represent ambiguous bases. To handle this case, LASTZ provides the ‑‑ambiguous=n option, which causes substitutions with N to be scored as zero. Additionally, the ‑‑ambiguous=iupac option causes the other IUPAC-IUB ambiguity codes (B, D, H, K, M, R, S, V, W, and Y) to be treated this same as an ambiguous N. The two ‑‑ambiguous options also allow you to specify rewards and penalties for matches and mismatches involving ambiguous characters.

In either case, non-ACGT characters are ignored during the seeding stage. Only seed words that consist entirely of A, C, G, and/or T are involved in seeding, even if the non-ACGT characters occur in "don't-care" positions in the seed pattern.

The score values described above can be changed if a scoring file is specified. The −1000 score is called bad_score and the −100 score is called fill_score. Further, which character is considered "bad" (by default this is X) can also be specified in the scoring file, and can actually be different between the target and query. Throughout this document, when we refer to the character X appearing in a DNA sequence, we generally mean the character specified as "bad", which defaults to X.

Splicing, or more correctly separation, can also be accomplished by placing a specific character between subsequences, then using the separator=<character> action. LASTZ will then break the sequence into the prescribed subsequences and prevent any alignment from crossing the boundaries.

Quantum DNA sequences are different: they use an arbitrary, user-defined alphabet of symbols, so the above-mentioned special treatments for N and X do not apply. The default "bad" character for quantum sequences is the null byte (00 hexadecimal), which is not even allowed in the sequence; however it can be changed to one of the valid alphabet symbols via the scoring file. There is no analog of ambiguous Ns for quantum sequences, as typically every symbol has some level of ambiguity.

Sequence Name Mangling

Often the names in the input sequence files are inconvenient for downstream processing, or create problems with certain output formats. This is further complicated by the fact that some input formats (most notably Nib) do not contain sequence names, so in those cases a name must be derived from the filename. LASTZ provides several choices for naming the input sequences. These alternatives are mutually exclusive; only one can be used at a time for a particular input file.

Internally, LASTZ handles the naming task in two phases. First, it creates a full header for the sequence. If the input format provides a name or header, that becomes the full header. Otherwise, the full header is constructed from the file name.

In the second phase, LASTZ shortens the full header to a nickname. If the full header starts with a file name, any path prefix is removed, and commonly-used file extension suffixes are also removed (.fa, .fasta, .nib, .2bit). Then by default, LASTZ uses the first word (composed of characters other than whitespace, vertical bar, or colon) of the remaining string as the sequence name. Thus a FASTA header like "> ~someuser/human/hg18/chr1.fa Human Chromosome 1" is shortened to simply chr1.

The actions nameparse=darkspace and nameparse=alphanum in the sequence specifier change how the first word is determined. darkspace (i.e., "non-whitespace") narrows the set of terminating characters to allow vertical bars and colons to appear in the word, while alphanum widens it so the word is restricted to only alphabetic, numeric, and underscore characters. Path prefixes and file extensions are still removed.

The default shortening is often adequate. For example, consider the following FASTA file. By default, the names will be 000007_3133_3729 and 000015_3231_1315.

    >000007_3133_3729 length=142 uaccno=FX9DQEU13H5YZN
    ACCCGAAAGAGAAACAGCTTCCCCCCCTGTCCCGAGGGATATCAAGTAGTTTGTTGGCTA
    GGCTGATATTGGGGCCTTCCGCTAGAGTCGGCGCCCGCGCCTACGAGTCCCCCCCACCCC
    CCACCCCCACAGCGGGTTATCC
    >000015_3231_1315 length=190 uaccno=FX9DQEU13HUTXE
    TTGTTGAGTCGGATGAGAATAGCAAGTGCAGTCAACGGCAATGTGCTGGGTTAGTACAAC
     ...

However, the user may find it more convenient to use the accession numbers. To accomplish this, she can use the nameparse=tag:uaccno= action. LASTZ will look for the tag string uaccno= in each header and read the name from the characters that follow it, up to the first character that is not alphabetic, numeric, or an underscore. In this case the sequence names would be FX9DQEU13H5YZN and FX9DQEU13HUTXE. If the tag string is not found in the full header for a particular sequence, the default shortening is used instead.

Now consider this FASTA file:

    >gi|197102135|ref|NM_001133512.1| Pongo abelii ...
    GCGCGCGTCTCCGTCAGTGTACCTTCTAGTCCCGCCATGGCCGCTCTCACCCGGGACCCC
    CAGTTCCAGAAGCTGCAGCAATGGTACCGCGAGCACGGCTCCGAGCTGAACCTGCGCCGC
     ...
    >gi|169213872|ref|XM_001716177.1| PREDICTED: Homo sapiens ...
    ATGTCTGAGGAGGTAGGATTTGATGCAGGAGGGAGGATCTGGTGCACTTATAAGGATCTG
    GGTCTGTCAGTGTCAGAGAAGGTAGGATCTGGCCCTGGTATGAGGATCTGGATCTGTCAG
     ...
    >gi|34784771|gb|BC006342.2| Homo sapiens ...
    GGGTGGGAGGACGCTACTCGCTGTGGTCGGCCATCGGACTCTCCATTGCCCTGCACGTGG
    GTTTTGACAACTTCGAGCAGCTGCTCTCGGGGGCTCACTGGATGGACCAGCACTTCCGCA
     ...

In this case the default action does not do what we want (all sequences would be named gi). The action nameparse="tag:gi|" gives us the names 197102135, 169213872, and 34784771. (Note the quotes; this is necessary to prevent the command-line shell from interpreting | as a pipe character.) Observe that a tag of ref| will not work, because the third sequence would need gb| instead.

Sometimes it is more convenient just to assign a specific name. This can be done with the nickname=<name> action. For example, using the target and query file specifiers ~someuser/human/hg18/chr1.nib[nickname=human] and ~someuser/human/ponAbe2/chr1.nib[nickname=orang], the output will show the sequences as human and orang rather than calling them both chr1. If <name> contains the substring {number}, the nickname will contain the number of the sequence within the file. This is particularly useful when there is more than one sequence in the file.

If you want to do away with name mangling entirely, you can use the action nameparse=full. This uses the full header as the sequence name. But note that if it contains spaces, the resulting alignment files may not be readable by downstream tools.

The above discussion applies to ordinary DNA sequences in FASTA, Nib, or 2Bit format. HSX index files are handled differently: by default LASTZ uses the name from the index as-is, without shortening it, and the various nameparse actions are not allowed. The nickname action can be used, but is generally not necessary since you can store the names you want directly in the index.

Note that if the subset=<names_file> action is used, the names in the <names_file> must match the mangled (or indexed) names.

For FASTA files, more complicated name mangling can be performed using standard Unix command-line tools. In the second example above, we could pipe the input through sed a couple times to shorten each name to the NCBI accession numbers NM_001133512.1, XM_001716177.1, and BC006342.2.

    cat query_file.fa \
      | sed "s/>.*ref\|/>/g" \
      | sed "s/>.*gb\|/>/g" \
      | lastz target /dev/stdin

Seed Patterns

Seeds are short near-matches between the target and query sequences, where "short" typically means less than 20 bp. Early alignment programs used exact matches (e.g. of length 12) as seeds, but spaced seeds can improve sensitivity when the sequences are diverged.

A spaced seed pattern is a list of positions, in a short word, where a seed may contain mismatches. For example, consider the seed pattern 1100101111. A 1 indicates a match is required in this position, and a 0 indicates a mismatch is allowed (effectively it is a "don't care" position). As the example below shows, using this seed pattern, the seed word GTAGCTTCAC hits twice in the sequence ACGTGACATCACACATGGCGACGTCGCTTCACTGG.

        target:  ACGTGACATCACACATGGCGACGTCGCTTCACTGG
    (mis)match:    ||XX|X||||          ||X|||||||
         query:    GTAGCTTCAC          GTAGCTTCAC
       pattern:    1100101111          1100101111

Spaced seeds have been shown to be more sensitive than exact match seeds, with little change in specificity. This is most advantageous when the sequences have lower similarity, such as human vs. mouse or chicken. Which seed pattern is best depends on the sequences being compared. See [Buhler 2003] for a discussion of spaced seeds and how to design them.

LASTZ’s seeding options give the “user” many choices. The intent is that these will be selected by some program (hence the quote marks around “user”), but they are available from the command line for anyone.

N-mer match:

    --seed=match<length>

General seed patterns:

    --seed=<pattern>

Half-weight seed patterns:

Single, double, or no transitions:

Filtering on transversions and matches:

    --seed=TTT0T00TT00T0T0TTTT --filter=2,15

Twin hit seeds:

    --twins=[<minsep>..]<maxsep>

Any-or-None Alignment

Sometimes, the only answer you want from an aligner is whether a query has any strong alignments to the target or not. For example, you may want to know which reads in a sequencing run have no alignment with a reference genome. In this case, if a read aligns to a thousand different places on a particular chromosome, you aren't interested in learning where — all you want to know is whether it aligned or not.

The ‑‑anyornone option is designed for such cases, and can significantly improve alignment speed. Once any qualifying alignment has been found, processing for the current query is halted. The alignment is reported to the output, and then we immediately begin processing the next query. A qualifying alignment is one that would normally be output given the other parameter settings; for example it satisfies the scoring thresholds (‑‑hspthresh and/or ‑‑gappedthresh) and any back-end filters.

To get a list of reads that have at least one "good" alignment with a reference sequence, you could do something like this:

    lastz <reference> <reads> --anyornone  \
      --step=10 --seed=match12 --notransition --exact=20 --noytrim \
      --match=1,5 --ambiguous=n \
      --filter=coverage:90 --filter=identity:95 \
      --format=general:name2

This option slightly changes the usual processing order described in the Overview. Instead of performing gap-free extension on all seeds, collecting them into a list of HSPs, and then performing gapped extension, each HSP is gap-extended and back-end filtered immediately. This avoids wasted work to perform complete early stage processing on hits that will just be abandoned as soon as the first qualifying alignment is found.

Y-drop Mismatch Shadow

The default configuration of gapped extension in LASTZ is to end the alignment where the score would be the highest. This means that any prefix or suffix of the alignment will have a non-negative score. While this is appropriate for alignments that lie somewhere in the middle of two long sequences, it is not desirable when an alignment is near the end of one or both sequences, which happens quite often when aligning short reads.

Consider the following alignment of a 50-base query to a chromosome target, and suppose we are using ‑‑match=1,5, ‑‑gap=6,1, ‑‑filter=identity:97, and ‑‑filter=coverage:95. The entire alignment as shown has 97.9% identity (46/47) and 100% coverage. However, the first five bases (AGAAC vs. AGAAG) have a negative score: four matches at +1 each and one mismatch at −5 gives a score of −1 for this prefix. The highest scoring alignment is from positions 6 through 50, for a score of 33 (the entire alignment scores only 32). If we stop the alignment at the highest score, coverage drops to 90%, and the alignment is discarded. The overall result is that we will discard reads that we don't want to, and we will see a bias against mismatches near the ends of reads. (Note that this anomaly arises because the alignment is terminated abruptly by the end of the sequence rather than normally by a low-scoring region; also the ‑‑filter=coverage option is more commonly used with short reads than with longer sequences.)

    target:  ... CTTAGAACGGTAGATACTTGTATAAT---CGAGGGGGTTATTTTGTACAAATGACT ...
                    ||||X||||||||||||||||||   ||||||||||||||||||||||||
     query:         AGAAGGGTAGATACTTGTATAATCAACGAGGGGGTTATTTTGTACAAATG
                         ↑                                           ↑
                    12345678901234567890123456789012345678901234567890

To avoid this behavior, use the ‑‑noytrim option when aligning short reads. This causes LASTZ to refrain from trimming such alignments back to the highest-scoring location. Specifically, if the gapped extension process encounters the end of the sequence, it will keep that as the end of the alignment. In this case a negatively-scoring prefix or suffix will be kept as long as it does not score worse than the ‑‑ydrop value.

Shingle Overlap

In some applications, e.g. when assembling reads into contigs, we want to determine how sequence ends overlap each other. For example, in case 1 below, the starting portion of the query overlaps the ending portion of the target by 30 bases, and both sequences extend beyond each other in opposite directions. We call this situation "shingling" (like shingles on a rooftop), and the shingle field of the General output format provides a measurement of it. A positive value indicates that the starting portion of the query overlaps the ending portion of the target (case 1), while a negative value indicates the roles are reversed (case 2). If neither of these cases occurs (e.g. if either sequence fails to extend beyond the other), an NA is reported.

Case 1 (shingle = +30):

                                                    target_end
                           3         2         1        ↓
                           098765432109876543210987654321
    target:  ... GACGGCGGCTAACACATTGTGTTGXACGTACCATAACCAA
                           ||||||X|||||||||XX||X||||||
     query:                AACACAGTGTGTTGCAACTATCATAACATTAAACTTTAGA ...
                           123456789012345678901234567890
                           ↑        1         2         3
                      query_start

Case 2 (shingle = −30):

                     target_start
                           ↓        1         2         3
                           123456789012345678901234567890
    target:                TCCCTAATAAATCTTAAGTGCGATCCGCAGCGAGGTGTAC ...
                              ||||X|||||||||X||||||||X||
     query:  ... TGGCGCCTGTAGTCTAAGAAATCTTAATTGCGATCCACAC
                           098765432109876543210987654321
                           3         2         1        ↑
                                                    query_end

Note that the value reported has no relation to the number of bases that align in that region, nor is it an indication that the alignment extends all the way to the start or end of the sequences. The shingle value is just evidence that the proper registration of the two reads is to overlap them by the given value — information that an assembler might use in assembling those reads into a contig.

Using Target Capsule Files

Target capsule files are provided to improve run-time memory utilization when multiple CPU cores on the same computer are running LASTZ with the same target sequence. They permit the lion’s share of the large internal data structures to be shared between the processes. This allows more copies of LASTZ to be run simultaneously with less physical memory, which can improve the throughput, for example, when mapping a large set of reads to a single (large) reference sequence.

To create a capsule file, use a command like this:

    lastz <target> --writecapsule=<capsule_file> [<seeding_options>]

To use the capsule file, run LASTZ like this:

    lastz --targetcapsule=<capsule_file> <query> [<other_options>]

Internally LASTZ asks the operating system to directly map the capsule file into the running program’s memory space in a read-only fashion. Multiple running instances can map the same file; each instance will have its own virtual addresses for the capsule data, but the physical memory is shared. There is no requirement for more than one instance to actually use the capsule simultaneously. Running a single copy of lastz with ‑‑targetcapsule will work fine, and in fact there may be a small speed improvement compared to running the same alignment without a capsule.

The downside of this technique is that the capsule files are very large and are also machine-dependent. For example, the file for human chromosome 1 is about 1.4 Gb. Note that attempts to run a capsule built on a mismatched computer are detected and rejected.

Inferring Score Sets

Scoring inference is an automated method for determining appropriate substitution scores and/or gap penalties directly from the sequences being aligned. The resulting scoring parameters can be saved to a file and/or used immediately to align the sequences. Generally these depend mostly on the species rather than particular regions, so once a suitable scoring set has been obtained for a pair of species, the inference does not need to be performed for each alignment run. In this section we give a brief overview of the inference process; see [Harris 2007] for a more detailed description.

Inference is achieved by computing the probability of each of the 18 different alignment events (gap open, gap extend, and 16 substitutions). These probabilities are estimated from alignments of the sequences. Of course, at first we don't have alignments, so we start by using a generic scoring set to create alignments, infer scores from those, then realign, and so on, until the scores stabilize or "converge". Ungapped alignments are performed until the substitution scores converge, then gapped alignments are performed (holding the substitution scores constant) until the gap penalties converge.

To have LASTZ infer scoring parameters, use a suitably enabled build of LASTZ (see below), and specify the ‑‑infer or ‑‑inferonly options. (The latter will stop after inferring the parameters, without performing the final alignment.) Settings for the inference process can be specified in a control file included with these options.

The ‑‑infscores option causes the inferred scoring parameters to be written out to a separate file. If no <output_file> is specified, it is written to the header of the alignment output file, as a comment. As a last resort, if no alignment is performed the scoring set is written to stdout. The parameters are written in the same format used to input scoring sets.

Usually it is undesirable to use all alignment blocks for inference. Blocks with a high substitution rate (low identity) are likely to be false positives. On the other hand, blocks with few substitutions (high identity) will be found regardless of what scoring parameters are used. Thus it is desirable to base the inference only on statistics from a mid-range of identity. By default the middle 50% is used (that is, the 25th through 75th percentile from the identity distribution), but this can be changed in the control file.

Special Builds Required:

Dynamic Programming Matrix

• Dynamic programming in general is a time-saving algorithm for computing values that can be expressed via a recurrence relation [Bellman 1957].
• It has long been used for affine gap alignments of DNA and protein sequences; see e.g. [Gusfield 1997].
• It uses a matrix of sequence positions to store partial results; early cells are used to compute later ones to avoid redundant work.
• Even for stages that do not involve gaps and do not actually use a DP algorithm, the matrix is helpful as a conceptual tool because of its strong correspondence with the dot-plot paradigm for visualizing sequence alignments.
• Here we use the convention of representing the target sequence from left to right along the horizontal axis (columns of the matrix), and the query sequence from bottom to top along the vertical axis (rows of the matrix); see e.g. Figure 5(a) for an example.
• Gap-free alignment segments lie along diagonals of the matrix/dotplot, in the forward (slope=+1) or reverse (slope=−1) orientation; the latter typically indicates an alignment on the reverse strand.
• Gaps bump the alignment to a different diagonal, since there is a progression in one sequence but not in the other; this has the effect of making gapped alignments look like diagonal-trending squiggly lines when drawn at low resolution.
• Diagonals are characterized by a constant difference between the target and query positions of their cells (or for a reverse diagonal, a constant sum).
• This matrix is not to be confused with the substitution scoring matrix, whose rows and columns correspond to characters rather than to sequence positions.

Filtering With Shell Commands

Though LASTZ provides several filtering options (e.g. ‑‑filter=identity, ‑‑filter=continuity, ‑‑filter=coverage, ‑‑filter=nmatch, ‑‑filter=nmismatch, ‑‑filter=ngap and ‑‑filter=cgap), sometimes these are not sufficient for the task at hand. But in many cases it is still possible to perform the desired filtering by using the ‑‑format=general option in conjunction with a simple awk, perl, or python script. Here we show one such example, using awk.

Suppose we want to filter alignments by length, discarding anything shorter than 500 bp, and that we need AXT output for downstream processing. We can have LASTZ output whatever columns are necessary to reproduce AXT and use awk to perform the filtering and reconstruct an AXT file.

Looking at the UCSC AXT specification, the corresponding --format=general fields are shown in the table below. Note that when determining which fields are needed for a given format, care has to be taken to make sure to get the correct start and end fields. Different formats count from zero instead of one, and some count reverse-strand positions along the plus strand. The interval coordinates section provides more detail about possible numbering schemes.

Then we can perform our filtered alignment with a series of commands like this:

  lastz target.fa query.fa \
     --format=general:name1,start1,end1,name2,start2,end2,strand2,score,text1,text2 \
   | grep -v "^#" \
   | awk '{ if ($3-$2+1 >= 500) print $0 }' \
   | awk 'BEGIN { n=-1;} { print ++n,$1,$2,$3,$4,$5,$6,$7,$8; print $9; print $10; print ""; }' \
   > filtered.axt

The first awk command computes the alignment length in the target, and if it is at least 500, copies the line to the output. $3 is end1 and $2 is start1. Since these represent a closed interval, we have to add 1 to get the length. $0 represents the entire input line.

The second awk command converts the alignment from a single line into four lines required for AXT. We use an awk counter, n, to create the alignment number field. The other fields are copied from the fields output by LASTZ.

Self-Masking a Sequence

For many alignment problems it is desirable to ignore alignments that consist soley of genomic repeats. For this reason, most finished genomic assemblies are soft-masked — bases that are part of identified repeats are in lowercase. LASTZ’s seeding stage avoids seed hits in lowercase, and thereby avoids finding alignments that are solely repeats.

With the wider availability of less-expensive DNA sequencing and custom assemblies, it is more common for users to have unannotated sequences. The following describes how LASTZ can be used to crudely identify duplications and soft-mask the original sequence. This process is called self-masking.

The self-masking process works by looking for alignments of the sequence with itself, and makes use of the dynamic masking feature to reduce computational time. The sequence is split into overlapping fragments, on the fly, which are then aligned against the entire sequence. As duplications are discovered they are marked as such and removed from the seeding process.

A command to perform self-masking would look like this, where critter.fa contains the sequence to be masked.

  cat critter.fa \
    | ../tools/fasta_fragments.py --fragment=200 --step=100 \
    | lastz critter.fa[multiple,unmask,nameparse=darkspace] /dev/stdin --masking=3 \
        --progress+masking=10K \
        --format=none --outputmasking+:soft=critter.masking.dat \
        --notransition

In that command, lastz is given the whole “critter” as its target sequence, and overlapping 200bp fragments of the critter as the queries.

The multiple action tells lastz to allow more than one sequence in the file, and the unmask action tells lastz to ignore any softmasking that may be present in the file. (If your sequence already has had some masking performed, and you want to keep that, omit unmask.) The nameparse=darkspace action tells lastz to extract the first non-whitespace string from the sequence header line. This is necessary to ensure that the final step (fasta_softmask_intervals) will see the same sequence names in the masked intervals file as those in the sequence file.

The ‑‑masking=3 option enables dynamic masking, which will mark any reference base appearing in 3 or more alignments. Since the fragments overlap by a factor of two, we expect every base will appear in two trivial alignments. Any more than that would be caused by a duplication elsewhere.

The ‑‑progress+masking option causes lastz to give you a progress report after every 10 thousand fragments. These reports come to the console (stderr) and look like this:

    (16.933s) processing query 50,001: critter_21299501, masked 8,920,893/51,304,566 (17.4%)

The ‑‑format=none option inhibits the normal alignment output and ‑‑format=outputmasking+:soft tells lastz to write the final masked intervals to a file.

The final line (‑‑notransition in this example) is whatever alignment scoring parameters you want to use. What is appropriate will depend on the level of divergence you want to allow in the masked duplications.

The command to apply the masking intervals to the fasta file will look like this:

  cat critter.fa \
    | ../tools/fasta_softmask_intervals.py --origin=1 critter.masking.dat \
    > critter.softmasked.fa

Aligning Many Subintervals

There are many occasions when you have a general idea of where the alignments you are interested in are, and it seems computationally wasteful to align entire sequences just to find a relatively few alignments. For instance, you may have identified some alignments using fast, high sensitivity settings and now want to look for alignments with higher divergence in the remaining regions. Or you may have previously found alignments but did not collect all the fields you needed. Or perhaps you used some tool other than LASTZ to identify regions where you want to focus your search. Or you may have gapped alignments from some other tool, and want to compare them to LASTZ's alignments in the same subsequences.

There are many ways to solve such a problem, and LASTZ provides several options to support these needs. Here we describe and critique several different approaches.

Sequence masking. LASTZ can use masking to eliminate the possibility of alignments in (or not in) a given list of intervals. So we can create two files, containing the desired intervals in the target and query, then run one LASTZ job.

The disadvantage of this solution is that you may get alignments between unintended interval pairs (for example, an alignment between the first target interval and the fifth query interval). Some post-proceesing would be necessary to eliminate these. Moreover, LASTZ will be spending a lot of its time looking for alignments in those unintended interval pairs.

Separate files. The simplest solution, conceptually, is to preprocess the sequences to extract the intervals of interest into separate files, then run each pair of files as a separate LASTZ job.

The disadvantages of this solution are numerous. There is extra I/O involved in splitting the files, and extra overhead in repeatedly launching LASTZ. Further, depending on your needs, post-processing may be necessary to map alignment positions back to the original sequences.

Subranges and subsets. The separate files solution can be improved upon by letting LASTZ mimic the file splitting, internally. This can be accomplished with the subrange and subset actions, while still running each as a separate lastz job.

This improves upon external file separation by eliminating some I/O, and eliminating the need to post-process to map positions. However, it still suffers the extra overhead of repeatedly launching LASTZ.

Alignment chores. Another solution is to use an alignment chores file (specified with the ‑‑chores=<file> option). The chores file corresponds directly to the interval pairs of interest. Complete alignment is performed over the regions defined by each pair (subject to whatever other options have been set), and is blocked from extending beyond the region.

There is little downside to this solution. The reported results will be the same as for the post-processed separate files solution, or subrange/subset solution (but with minor variations such as shifting of equally-scoring gap placements). There is some computational waste in the chores solution, but this is much less costly than the repeated launching.

In most cases, a chores file will be preferred over an anchors file.

Anchor segments. Another solution is to use an anchor segments file (specified with the ‑‑segments=<file> option). The anchors file does not correspond directly to the interval pairs of interest. Instead, you will need to have same-length intervals in target and query. Typically this will be a single point somewhere in the region. Alignment is anchored at this point. Gapped alignment is performed in both directions from that point, and is not restricted to the region.

The main disadvantage of this solution is lack of versatility. It is most appropriate when used in conjunction with an external anchor-identifying method.

It can also be useful in cases where you want to run ungapped alignment with a different scoring scheme than for gapped alignment. An ungapped run of LASTZ creates the anchors (segments) file, and a second run uses those as anchors for ungapped alignment with a different scoring scheme.

Differences from BLASTZ

• BLASTZ had a "short-by-two" error, which has been corrected in LASTZ. In many cases, BLASTZ shortened alignments by two bases on either or both ends.

• BLASTZ had a problem with premature alignment termination; this has been corrected in LASTZ.

• BLASTZ used the ydrop value from the main alignment as the xdrop value for interpolation; this has been corrected in LASTZ.

• BLASTZ had a problem when ydrop is less than the penalty for a one-base gap; this has been corrected in LASTZ.

• BLASTZ chaining had a problem that caused it to discard very high-scoring HSPs; this has been corrected in LASTZ.

• The handling of ties in the DP matrix was unspecified in BLASTZ. This has changed in LASTZ, which specifically prefers a longer alignment to a shorter alignment with the same score. This change reflects the use of LASTZ to align short reads, and desire to align as much of the read as possible.

• The handling of bounding alignments in the DP matrix is different in LASTZ than in BLASTZ. This is discussed in Bounding Alignments in the DP Matrix. The ‑‑allgappedbounds option can be used to revert to the bounding criteria used in BLASTZ.

• The handling of amibiguous nucleotides has been clarified in LASTZ, and in some cases the default behavior is different than in BLASTZ. By default, BLASTZ allowed IUPAC-IUB ambiguity codes (B, D, H, K, M, R, S, V, W, and Y) in fasta sequences but was unclear about how these were scored. Since we feel the user should be aware of how these bases are treated, LASTZ rejects them by default. The ‑‑ambiguous=iupac option permits them but treats them the same as an ambiguous N. This is discussed in Non-ACGT Characters.

• LASTZ can produce a variety of alignment output formats such as AXT, MAF, and human-readable text, as well as BLASTZ’s LAV format.

• LASTZ can take the guesswork out of selecting alignment scoring parameters by inferring them for you, based on its analysis of the input sequences.

• LASTZ provides a large variety of seeding options.

Bounding Alignments in the DP Matrix

During the gapped extension stage, LASTZ processes the anchors in order of score (highest scoring anchor is extended first, and so on). As anchors are extended, a list of bounding alignments is constructed. These correspond to paths in the DP matrix. Bounding alignments created for higher-scoring anchors are used to bound the possible DP paths that lower-scoring anchors can take. This prevents alignments from crossing each other.

In BLASTZ, every gapped extension became a bound, and this was originally the default behavior in LASTZ, through release 1.1.52. However, this caused LASTZ to miss some alignments which it should have found. The failure case occured as follows. A high-scoring anchor is extended but fails to meet the score threshold. But it gets added as a bound. Then the extension of a lower-scoring anchor is prevented from crossing or intersecting with that path, and it too gets discarded even though it might score highly enough. This could occur, for example, when two extensions would (in the absence of each other) share the same tail, and the higher-scoring of the two has a lower-scoring anchor.

The correction for this is to only use alignments as bounds if they satisfy the score threshold. This corrected behavior is now the default in LASTZ (as of release 1.02.00). The ‑‑allgappedbounds option can be used to revert to the bounding criteria used in BLASTZ.

Change History

    quantum.c: In function 'generate_dna_ball':
    quantum.c:347: error: array subscript is above array bounds

  WARNING. Scores file may warrant setting of thresholds absent from scores.txt.
  Minimum match score is 10, for matrix entry (A,A).
  This may not work well with default --gappedthresh=3000.

References

Bellman R (1957). Dynamic Programming. Princeton University Press, Princeton, NJ.

Buhler J, Keich U, Sun Y (2003). Designing seeds for similarity search in genomic DNA. Proc. 7th Annual International Conference on Research in Computational Molecular Biology (RECOMB '03), pp. 67-75.

Chiaromonte F, Yap VB, Miller W (2002). Scoring pairwise genomic sequence alignments. Pacific Symposium on Biocomputing 7:115-126.

Cock PJA, Fields CJ, Goto N, Heuer ML, Rice PM (2009). The Sanger FASTQ file format for sequences with quality scores, and the Solexa/Illumina FASTQ variants. Nucleic Acids Research 38:1767-1771.

Gusfield D (1997). Algorithms on strings, trees and sequences. Cambridge University Press, Cambridge, pp. 244.

Harris RS (2007). Improved pairwise alignment of genomic DNA. Ph.D. thesis, Pennsylvania State University.

Li H et al. (2009). The Sequence Alignment/Map (SAM) format and SAMtools. Bioinformatics 25:2078-2079.

Myers EW, Miller W (1989). Approximate matching of regular expressions. Bull. Math. Biol. 51:5-37.

Zhang Z, Berman P, Miller W (1998). Alignments without low-scoring regions. J. Comput. Biol. 5:197-210.

Acknowledgments

Thanks for to Haibao Tang for contributing an example implemention for BLASTN output.