Make insert size thresholds robust and thread safe

[jrobinso]

Review of the insert size (TLEN) threshold estimate in PEStats, prompted by igvteam/igv.js@96e3776, which fixed the equivalent code in igv.js. The load-order bug fixed there is not present here — desktop has always pooled samples across loads and recomputed — but the review turned up two other problems.

Thread safety

track.load(frame) runs on a 5 thread pool, so in multi-locus view several tiles of the same track load at once, all writing the shared peStats map and the shared PEStats objects with no synchronization. Concurrent DoubleArrayList.add could lose a sample or leave an uninitialized 0.0 in the pool (enough zeros drag the 0.5th percentile to 0 and silently disable small-insert coloring), and toArray()'s trimToSize() could swap the backing array under a concurrent add, throwing ArrayIndexOutOfBoundsException or NumberIsTooLargeException out of loadTile — an error dialog, and a tile returned before finish().

Each tile now samples into its own thread-confined PEStats (PEStats.forLoad) and merges once when the load completes. merge, computeInsertSize and computeExpectedOrientation are synchronized; the published thresholds are volatile so the renderer reads them without a monitor. The shared map is a ConcurrentHashMap.

Robustness of the estimate

The thresholds are percentiles in the tails, so a region dense with proper-pair-flagged discordant reads could drag the maximum past every read worth flagging. Following igv.js:

• outliers more than 10 robust standard deviations (median absolute deviation) from the median are trimmed before the percentiles are taken, unless that would discard a majority of the sample
• the pool grows from 1000 to 20 000 samples, and no single load may contribute more than 1000, so one deep region can no longer own it

Also

• loadTile recomputed the thresholds from the global preferences on every load, discarding percentiles set per-track through Set insert size options .... It now uses the track's RenderOptions, which fall back to the same preferences when unset.
• A minimum percentile of 0 threw OutOfRangeException out of the load. It now means what it reads as: no read is colored as a small insert. The dialog accepts 0 accordingly.
• minTLENPercentile / maxTLENPercentile were written to JSON sessions as numbers but read back with getString, so a session carrying them failed to load. igv.js's legacy minFragmentLength / maxFragmentLength are now accepted as aliases for minTLEN / maxTLEN.
• InsertSizeSettingsDialog restored a Swing component dump instead of the previous value after a bad entry.

Notes

Insert size coloring is not a useful metric for RNA data, so no accommodation is made for libraries whose large TLENs are genuine (intron-spanning pairs); desktop and igv.js behave identically here.

Three tests in PEStatsTest cover the trimming, the per-load cap and the zero minimum; each was verified to fail without its fix. Full suite: 681 tests, 0 failures. User documentation for the estimator is drafted separately for igv.org.

🤖 Generated with Claude Code