GitLab

I've verified that freebayes now only works using whole-chromosome sequences.
The anti-pattern of trying to cache little bits of the reference has been
resolved by removing offending methods in the AlleleParser. The system should
also run much faster.

This likely affects #163, #166, #168, #170, but reporters need to test before I
close these.

Fix link to Brad's blog post.

It's much easier to just load the entire sequence we're working on. This commit
does that.

Avoids an infinite loop.

Only try to step to the next place with data when we've *passed* the last of
the current alignments.

And the tests pass...

A previous edit had removed an important call to
loadReferenceSequence(alignment) when using a target region. This
yielded #155.

Should work by default, the only dep is cmake, which tends to be on travis.

To add tests, drop them into test/t/. bash-tap is used, and included here.

The idea is to use command-line tests of the function of the method.

A tiny test region from NA12878 is included.

For a long time, the best way to genotype diploids using freebayes was by
specifying --experimental-gls.  This model incorporates a lot of things into
the genotype likelihoods (GLs) that matter, such as the effective base depth
metric (Q = MQ * BQ) from SNPTools, and partial observation support (which
matters for big deletions, and long haplotypes).

However, bugs when handling polyploids prevented me from setting it as default.

With this commit, these issues are resolved and the default is now set to what
was previously specified by "--experimental-gls".

Use --legacy-gls for old polybayes-equivalent operation.

After rewrite it was omitted.

It was going to add a mb per compile! No way!

Resolves a bug around long indel detection. This was related to the change of
underlying Allele objects between when partial support was established and when
it was used (phew).

Basic automated tests for region/target handling

Region without coords

This uses the Test-Simple bash library to support TAP output.  TAP
output means that it works with pretty much any testing framework.
prove is a common way to run tests:

    prove -v test/

They used stopPos = -1 to signal "until the end of the reference", but
the reference length (1-origin, inclusive) was used directly instead of
also being converted to a 0-origin, inclusive bound.  Other targets like
chrZ:0-200 are already properly adjusted to be fully-closed.

This reverts commit 22295cfc.

The right fix for this is actually a few lines above:

    (stopPos == -1) ? reference.sequenceLength(startSeq) : stopPos - 1); // internally, we use 0-base inclusive end

should be

    (stopPos == -1) ? reference.sequenceLength(startSeq) - 1 : stopPos - 1); // internally, we use 0-base inclusive end
or
    ((stopPos == -1) ? reference.sequenceLength(startSeq) : stopPos) - 1); // internally, we use 0-base inclusive end

and then this commit should be reverted. Other targets like chrZ:0-200
are already adjusted to be fully-closed in targets, so the correct
comparison is bd.right + 1 > reference.sequenceLength(...). The bug is
that targets like chrZ were left half-open by mistake.

I missed this when I was auditing!