last-split
==========

This program estimates "split alignments" (typically for DNA) or
"spliced alignments" (typically for RNA).

It reads candidate alignments of query sequences to a genome, and
looks for a unique best alignment for each part of each query.  It
allows different parts of one query to match different parts of the
genome.  This is useful for DNA queries that cross rearrangement
breakpoints, or RNA queries that cross splice junctions.

Examples
--------

Split alignment of DNA reads to a genome
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Assume the DNA reads are in a file called "q.fastq" (in fastq-sanger
format), and the genome is in "genome.fasta" (in fasta format).  We
can do the alignment like this::

  lastdb -m1111110 db genome.fasta
  lastal -Q1 -e120 db q.fastq | last-split > out.maf

Spliced alignment of RNA reads to a genome
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Now we assume that "q.fastq" has reads from RNA forward strands.  This
time, we provide the genome information to last-split, which causes it
to do spliced instead of split alignment, and also tells it where the
splice signals are (GT, AG, etc)::

  lastdb -m1111110 db genome.fasta
  lastal -Q1 -e120 db q.fastq | last-split -g db > out.maf

This will favour splices starting at GT (and to a lesser extent GC and
AT), and ending at AG (and to a lesser extent AC).  However, it allows
splices starting and ending anywhere.  It also favours splices with
introns of typical length, specified by a log-normal distribution
(i.e. cis-splices).  However, it allows arbitrary trans-splices
between any two places in the genome.  If you wish to turn off
trans-splicing, add -t0 to the last-split options.

Unfortunately, the cis-splicing calculations can be slow, depending on
various factors such as query sequence length.  If it is too slow,
please try the next recipe.

Faster trans-spliced alignment
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Here we use -c0 to turn off cis-splicing, and -t0.004 to specify a
higher probability of trans-splicing::

  lastdb -m1111110 db genome.fasta
  lastal -Q1 -e120 db q.fastq | last-split -c0 -t0.004 -g db > out.maf

FAQ
---

:Q: Before aligning RNA, should poly-A tails be trimmed?

:A: It's not essential, but it might make things faster.  Poly-A
    tracts tend to have many matches in the genome.  By trimming, you
    can prevent lastal and last-split from wasting time on such
    matches.

Going faster by parallelization
-------------------------------

For example, split alignment of DNA reads to a genome::

  parallel-fastq "lastal -Q1 -e120 db | last-split" < q.fastq > out.maf

This requires GNU parallel to be installed
(http://www.gnu.org/software/parallel/).

Output
------

The output is in MAF(-like) format::

  a score=150 mismap=0.000413
  s chr21  15963638 25 + 48129895 TCAGATGAGGACCTAATTTATTACT
  s query7       50 25 +       75 TCAGATGAGGACCTAATTTATTACT
  q query7                        EBEEC@CE=EEE?FEDAED5?@@D@
  p                               !#$'BBBBBBBBBBBBBBBBBBBBB

The "mismap" is the estimated probability that this part of the query
should be aligned to a different part of the genome.  The line
starting with "p" indicates the probability that each base should be
aligned to a different part of the genome.  It uses a compact code:

  ======  =================   ======  =================
  Symbol  Error probability   Symbol  Error probability
  ------  -----------------   ------  -----------------
  ``!``   0.79 -- 1           ``0``   0.025 -- 0.032
  ``"``   0.63 -- 0.79        ``1``   0.02  -- 0.025
  ``#``   0.5  -- 0.63        ``2``   0.016 -- 0.02
  ``$``   0.4  -- 0.5         ``3``   0.013 -- 0.016
  ``%``   0.32 -- 0.4         ``4``   0.01  -- 0.013
  ``&``   0.25 -- 0.32        ``5``   0.0079 -- 0.01
  ``'``   0.2  -- 0.25        ``6``   0.0063 -- 0.0079
  ``(``   0.16 -- 0.2         ``7``   0.005  -- 0.0063
  ``)``   0.13 -- 0.16        ``8``   0.004  -- 0.005
  ``*``   0.1  -- 0.13        ``9``   0.0032 -- 0.004
  ``+``   0.079 -- 0.1        ``:``   0.0025 -- 0.0032
  ``,``   0.063 -- 0.079      ``;``   0.002  -- 0.0025
  ``-``   0.05  -- 0.063      ``<``   0.0016 -- 0.002
  ``.``   0.04  -- 0.05       ``=``   0.0013 -- 0.0016
  ``/``   0.032 -- 0.04       ``>``   0.001  -- 0.0013
  ======  =================   ======  =================

Other symbols indicate lower error probabilities, and "~" is the
lowest possible.  In general::

  Error probability <= 10 ^ -((ASCII value - 33) / 10)

The "mismap" is simply the lowest probability from the "p" line.

Split versus spliced alignment
------------------------------

Here is a split alignment::

  Query         ttctttgat--gctagtcctgatgttatggtattttttatcgaatgataa
                  |||||||--||||||                |||x||||||||||||
  Genome chrA  ...ctttgatatgctagt...             |||x||||||||||||
  Genome chrB                                 ...tttatatcgaatgata...

And here is a spliced alignment::

  Query        ctagtcgatatt--gctgtacgtctgttagctat-tttttcctctgtttg
                  |||x|||||--|||||||||----|||||||-|||||x|||||
  Genome chrA  ...gtctatattatgctgtacgt... |||||||-|||||x|||||
  Genome chrB                          ...tagctatattttttctctg...

Split alignment allows arbitrarily large unaligned parts in the middle
of the query, whereas spliced alignment applies a standard gap
penalty.  (Both allow arbitrarily large unaligned parts at the edges
of the query.)

For DNA queries, you might wish to try "trans-spliced" alignment
without considering splice signals::

  lastdb -m1111110 db genome.fasta
  lastal -Q1 -e120 db q.fastq | last-split -c0 > out.maf

Options
-------

  -h, --help
         Show a help message, with default option values, and exit.

  -g, --genome=NAME
         Do spliced alignment, and read splice signals (GT, AG, etc)
         from the named genome.  NAME should be the name of a lastdb
         database.

  -d, --direction=D
         Do spliced alignment, and set the strandedness of the
         queries: 0=reverse, 1=forward, 2=unknown/mixed.  This
         determines whether forward and/or reverse-complement splice
         signals are used.

  -c, --cis=PROB
         Do spliced alignment, and set the average probability per
         base of cis-splicing.  The default value roughly fits human
         RNA.

  -t, --trans=PROB
         Do spliced alignment, and set the average probability per
         base of trans-splicing.

  -M, --mean=MEAN
         Do spliced alignment, and set the mean of ln(intron length).
         The default value fits human RNA.

  -S, --sdev=SDEV
         Do spliced alignment, and set the standard deviation of
         ln(intron length).  The default value fits human RNA.

  -m, --mismap=PROB
         Don't write alignments with mismap probability > PROB.
         Low-confidence alignments will be discarded unless you
         increase this value!

  -s, --score=INT
         Don't write alignments with score < INT.  The default value
         is somewhat higher than the lastal score threshold.
         Specifically, it is e + t * ln(1000), where e is the score
         threshold, and t is a scale factor that is written in the
         lastal header.  This roughly means that, for every alignment
         it writes, it has considered alternative alignments with
         one-thousandth the probability.

  -n, --no-split
         Do probability calculations as usual, but write the
         *original* alignments, annotated with "p" lines and mismap
         probabilities.  Note that the mismap and score limits still
         apply.

  -v, --verbose
         Show progress information on the screen.

  -V, --version
         Show version information and exit.

Details
-------

* The input must be in MAF format, and it must include header lines
  (of the kind produced by lastal) describing the alignment score
  parameters.

* The program reads one batch of alignments at a time (by looking for
  lines starting with "# batch").  If the batches are huge
  (e.g. because there are no lines starting with "# batch"), it might
  need too much memory.

* lastal can optionally write "p" lines, indicating the probability
  that each base is misaligned due to wrong gap placement.
  last-split, on the other hand, writes "p" lines indicating the
  probability that each base is aligned to the wrong genomic locus.
  You can combine both sources of error (roughly) by taking the
  maximum of the two error probabilities for each base.

The following points matter only if you are doing something unusual
(e.g. bisulfite alignment):

* If the header has more than one score matrix, last-split will use
  the first one.

* It assumes this score matrix applies to all alignments, when the
  alignments are oriented to use the forward strand of the query.

Limitations
-----------

last-split does not support:

* Different scores for insertions and deletions.
* Generalized affine gap costs.

To do
-----

* An option to specify splice signals and their strengths.
