pirl-unc / hitlist / 34300209285
84%

DEFAULT BRANCH: main
Ran
Jobs 1
Files 30
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 83.721% (+0.08%) from 83.644%
34300209285

push

github

web-flow 
Resolve the sample system before the arm, via sample_group (#359, #364) (#446)

The arm scorer reads sample_label and perturbation as one bag of tokens, so
it cannot tell a system descriptor from a condition one. Two consequences,
both live:

Extra identifying words on one arm of a pair decided the pair. PMID
29242379's untreated UWB arm alone carried "(ovarian carcinoma)", and
`ovarian` matching source_tissue "Ovary" handed it all 3,919 UWB rows with
nothing about treatment in evidence; PMID 30833945 lost 4,676 rows the same
way via `lung`. That is 8,595 rows confidently attributed to a control arm
on no evidence.

And because the narrative fields are blocked wholesale whenever arms
disagree, a system whose arm is unambiguous was not attributable either --
29242379's TIL and meningioma systems have one arm each and reached nothing.

sample_group splits the two questions. The system stage admits the narrative
fields, because naming a system is what they do reliably and they say
nothing about treatment; the arm stage keeps blocking them. Opt-in per
study, enforced all-or-nothing at load so a half-curated study cannot
silently fall back to the ungrouped path. When arms of a known system tie,
the row reports the system and withholds the arm as group_ambiguous, which
required adding to the class-pool path the _consensus_meta fallback the
allele path already had. Symmetric arm labels are now a tested invariant:
every arm of a multi-arm group must contain its group name.

Curated four studies:

- 29242379 Chong: four systems, two of them treated/untreated pairs. TIL
  (11,012) and meningioma (438) now attribute to their arm; B-LCL (66,959)
  and UWB (3,919) report the system.
- 30833945 Javitt: one system, two arms, no per-row arm evidence anywhere in
  the study; all 7,524 rows now report the system instead of 4,676 of them
  naming the wrong arm.
- 32938616 Faridi: three cell lines, not a panel. All 92,169 rows resolve to
  a line; 37,643 keep their condi... (continued)

6953 of 8305 relevant lines covered (83.72%)

0.84 hits per line

Coverage Regressions

Lines Coverage ∆ File
105
91.0
 0.25% export.py
49
94.48
 0.03% curation.py
Jobs
ID Job ID Ran Files Coverage
1 34300209285.1 30
83.72
 GitHub Action Run
Source Files on build 34300209285