SAMBLASTER duplicate criteria

[mattmaurano]

SAMBLASTER duplicate criteria: Within a group of read to be marked duplicate, how does samblaster pick the one to keep? Does it consider base qualities, read length, num mismatches to reference, etc.?

Alternatives:
MarkDuplicates (not Cigar one) -- does sorting on readname fix past memory usage issues? Does it still mark supplementals in that case?
https://broadinstitute.github.io/picard/command-line-overview.html#MarkDuplicatesWithMateCigar

bamUtil

sambamba -- FrickTobias/BLR#113

[andremrsantos]

SAMBLASTER uses a single-pass approach to improve speed, thus he always keeps the first read and marks all other as duplicates without considering any other criteria. They claim that due to Illumina high-quality this strategy makes little difference in practice.

Here is their quote on the supplemental material:

However, two-pass algorithms have the advantage that they can choose to keep from amongst a set of duplicates, the read pair with the “best” score using some metric that differentiates sequence and/or alignment quality. In contrast, SAMBLASTER’s one pass approach can only keep the first pair from a set of duplicate pairs. Given the high quality of Illumina paired-end sequencing, we find this makes little difference in practice.