overlap assembly and coverage trimming to create VariantSequence candidates

[iskandr]

The primary change in this branch is that I added a variant_cdna_sequence_assembly option to the call chain which determines the cDNA sequence from reads (and a corresponding --variant-sequence-assembly option to the CLI). I also added coverage statistics and trimming by coverage to VariantSequence.

Less important changes:

• Reorganized ReferenceContext by moving the namedtuples from which it inherits fields into separate modules and renaming them to ReferenceSequenceKey and ReferenceCodingSequenceKey.
• Turned simple functions for constructing objects (e.g. allele read from locus read) into class methods.
• Caching PyEnsembl on Travis

TODO for a future branch:

• If we can't establish a ReferenceContext for a variant sequence with only 1 reading supporting the very beginning of its sequence, try trimming to the next highest coverage level(s) until a reference context can be established. (Trim VariantSequences by increasing coverage while trying to figure out reading frame #58)

• merge overlapping paired reads into single LocusRead objects (extends the length of sequences we can construct). (Merge overlapping read pairs into single LocusRead #59)

This change is

[coveralls]

Coverage decreased (-0.5%) to 86.411% when pulling dcfd2ed on use-assembly-and-trimming into f60e654 on master.

[ihodes]

LGTM, would love to go through this with you a bit, and @julia326 is going to look through this as well.

[ihodes]

Can't comment on the actual line, but curious what the reasoning is behind a default min_overlap_size of 30?

[ihodes]

Additionally, it should probably be set to MINIMUM_OVERLAP_SIZE?

[ihodes]

Otherwise, I reviewed this entire module, and it looks good to me!

[iskandr]

@ihodes I'm worried about "false" assemblies that only overlap on the variant nucleotides (but actually come from different splice isoforms). It seems that requiring some amount of overlap beyond the variant nucleotides will decrease the chances of false assembly but the exact amount (30) is pretty arbitrarily chosen: shorter than a read length but long enough to be a kmer unlikely to happen by chance.

[ihodes]

Should MIN_OVERLAPPING_READS go here as well? What makes it into here?