Improve amplicon detector

The current technique that the assembled version is using to detect amplicons has two problems:

1. Some lengths are identified as amplicons when they shouldn't be.
2. Some lengths are not identified as amplicons when they should be. Usually, this is because there's another amplicon of a similar length, and it raises the average count within the 20 length window.

The 12 Jun 2015 run has examples of both problems.

• V3LOOP amplicon not detected, because it didn't quite outnumber its neighbours enough.
  • 57368A-10-HLA-B-63594A-V3-1-V3LOOP_S10
  • 57481A-2-HLA-B-63456A-V3-2-V3LOOP_S14
  • 57481A-3-HLA-B-63456A-V3-3-V3LOOP_S15
• HLA exon3 amplicon too short, because the full-length one got squished between two other spikes.
  • 57482A-2-HLA-B-63461A-V3-2-V3LOOP_S26

Instead of choosing the threshold for amplicon peaks as a multiple of the average count, use a multiple of the 90th percentile. There are several parameters we could try to optimize:

• window size to compare each peak to, currently +/- 20

• percentile to compare to within the window, currently 90

• multiple of comparison, currently 50

• minimum count, currently 50

• change from average to percentile, and choose reasonable thresholds.

• optimize the thresholds over a large number of samples.

No files or tests are named. Start by locating the assembled amplicon detector and reproducing the 12 Jun 2015 examples, then compare its peak thresholds with the listed parameters. Done means correcting the V3LOOP and HLA exon3 detections and evaluating threshold choices across a large sample set.