clintval / vartovcf / 39539207865

98%

push

github

feat!: sort output records by coordinate with spillsort (#148)

* feat: fold leaky MNV partials back into their MNV with --fix-leaky-mnv

VarDict writes the reads of an MNV that end, have a no-call or a low-quality base inside it as shorter SNV or MNV rows. With --fix-leaky-mnv these partial rows are grouped with their MNV by UMI-tools' directional adjacency rule (parent >= 2 * child - 1 in every sample) within each VarDict region and folded into it the way VarDict's own MNV merging folds reads, keeping AD[1] within DP.

* docs: call out the leaky-MNV fold's effect and VarDict's own MNV fold

* docs: shorten the Leaky MNVs section and keep only averages from real data

* docs: cite only UMI-tools for the adjacency rule

* docs: explain how a partial two MNVs can claim is assigned

* docs: keep only what a user needs about the fold's limits

* fix: let an MNV no root has room for take its own partials

An MNV that a larger MNV could claim by reads but not fit beside its first-base reads was written alone and could not take partials of its own; it now becomes a root. The module doc also states VarDict's left fold as it is: a first-base partial whose reads end, on average, within it.

* perf: walk the two-sided Fisher test out from the most likely table

Each table's probability comes from its neighbour's by one ratio, so only the most likely table needs log-gamma terms, and each tail stops below e^-40 of the observed table. It matches the previous sum to 3e-11 relative on 186,000 random tables and is about 4 to 300 times faster.

* docs: describe the fold's DP and read means as computed, and test FILTER on merged counts

DP adds a later partial's reads as VarDict's own fold does, the read means are weighted means of VarDict's rounded row values, and the README says folding never pushes AD[1] above DP rather than that a folded record never has it there. A tumor-normal test shows a fold moving an MNV from LOW_AF to PASS.

* feat!: normalize every record a... (continued)

509 of 534 new or added lines in 4 files covered. (95.32%)

3174 of 3236 relevant lines covered (98.08%)

217492.43 hits per line