jinyang-omentum-subtype-artifact-mechanism-canary-expression-matrix-v1

Canary QC artifact for the expression arm of jinyang-omentum-subtype-artifact-mechanism. Shiba v0.8.2 expression matrices for 40 canary samples (20 NovaSeq / 20 non-NovaSeq) over 78,724 genes, from STAR 2nd-pass BAMs against the Ensembl 113 annotation.

This is canary-scale QC data, not a result. 40 of the cohort's 160 samples, on a 78,724-gene matrix. It exists to prove the pipeline runs, to measure what it costs, and to expose anything wrong before the full arm is submitted.

What was run

Stock Shiba src/expression.py with one deliberate deviation: featureCounts is invoked stranded (-s 2, reverse / fr-firststrand) via a bind-mounted patch over /opt/Shiba/src/expression.py. Upstream Shiba passes no -s at all, i.e. unstranded. The bind mount is gated (the job greps the mounted file for the flag and aborts if it did not take effect), because a silent bind failure would produce a plausible-looking unstranded matrix under a stranded label.

The strandedness choice rests on the s1/s2 asymmetry, not on a comparison against s0. Measured on one canary sample restricted to chr21+chr22 (10,488,208 reads):

• -s 0 assigned 0.2626 (NoFeatures 1,937,948)
• -s 1 assigned 0.0473
• -s 2 assigned 0.2485

s2/s1 = 5.25x. That forward/reverse split is the signature of a reverse-stranded library. Note that -s 2 assigns less than -s 0 (0.2485 vs 0.2626) — the criterion originally written into the sbatch ("-s 2 should be the highest of the three") does not hold, and it was dropped in favour of the asymmetry. The full control is a separate artifact: jinyang-omentum-subtype-artifact-mechanism-canary-strandedness-control-v1.

Validity gates: 9/9 PASS

Gate	Status	Detail
sample_columns_match_input	PASS	40 TPM columns vs 40 requested samples
gene_count_matches_gtf	PASS	78724 genes in the matrix vs 78724 gene features in homo_sapiens.gtf
no_nan_or_inf	PASS	0 non-finite cells
non_negative	PASS	0 negative cells
no_empty_sample_columns	PASS	zero-valued columns: none
tpm_columns_sum_to_1e6	PASS	max
counts_are_integers	PASS	all count cells integral
library_sizes_positive	PASS	library size range 34,086,598 - 77,908,511
bh_fdr_matches_statsmodels	PASS	hand-rolled BH vs statsmodels fdr_bh on the same 2000 p-values; max

Two of these gates failed on the first run, and both failures were bugs in the gates

Worth recording, because a reader who finds the earlier job log will otherwise read those failures as data problems. They were not.

1. gene_count_plausible failed at 78,724 genes against a hard-coded range of (10,000, 70,000). The ceiling had been set from a recollection of the protein-coding-plus-lncRNA count rather than from the annotation actually in use. Measured directly from the pinned GTF: 78,724 gene features, 78,724 unique gene_ids, zero version suffixes — exactly the matrix's row count. The data was right; the threshold was wrong. Fixed by widening the range and by adding an exact check (gene_count_matches_gtf) that counts the GTF's gene features and requires equality, which cannot rot when the annotation changes.
2. bh_fdr_matches_statsmodels compared Benjamini-Hochberg adjustment over the full gene set against statsmodels run on only the first 2,000 p-values. BH is a function of the whole vector — both its length and each p-value's rank — so the two can only agree when the vector happens to be exactly 2,000 long. bh_fdr itself was correct throughout (it matches an independent implementation at 0.000e+00 for n = 2,000 / 20,000 / 78,724). Fixed by comparing both on the same vector; the rerun gives max |diff| = 1.11e-16.

Throughput: the number this canary existed to produce

The red-team brief flagged the full-run cost as an untested 90-160 min extrapolation.

• per-sample median 90.0s (range 64-461s) over 40 samples
• measured total wall time for 40 samples: 5537s (92.3 min)
• extrapolated to the full 160-sample arm: 4.00h (median) / 6.18h (mean), single process

So the brief's estimate was optimistic by roughly 1.5-2.5x, and the full arm is a 4-6 hour job, not a 1.5-2.5 hour one. That matters for whether the full run needs batching to survive an 8 h wall clock.

Housekeeping drift between arms (reported, not gated)

gene	Ensembl	median TPM	CV	arm p
ACTB	ENSG00000075624	2846.4	0.602	0.925
GAPDH	ENSG00000111640	2808.0	0.682	0.0179
TUBB	ENSG00000196230	651.3	0.398	0.036
RPL13A	ENSG00000142541	1251.1	0.454	0.695
B2M	ENSG00000166710	1908.6	0.591	0.86
HPRT1	ENSG00000165704	43.5	0.837	2.69e-06

Housekeeping genes are not gated, deliberately: if two platforms really do differ systematically, GAPDH moving with them is a finding rather than a failure. But the red-team brief requires GAPDH and ACTB not be significant, and GAPDH is (p = 0.0179); HPRT1 is strongly so (p = 2.7e-06). ACTB, B2M, RPL13A and TUBB are not significant. Three of six differ at raw p < 0.05 where ~0.3 would be expected by chance. This is the clearest single signal in the matrix and it points at a systematic difference between the arms — see the composition warning below, which gives a non-platform explanation for it.

The arm contrast is not a clean platform contrast

NovaSeq vs non-NovaSeq at full cohort scale (n = 160; 113 vs 47) is entangled with several clinical and processing variables, all measured against clinical.tsv:

variable	NovaSeq	non-NovaSeq	test
seq_set	Sets 9a-18 (12 sets)	Sets 1-8b (12 sets)	perfectly disjoint, chi2 p = 8.3e-25
cohort	DECIDER 57, HERCULES 46, MUPET 7, OC_legacy 3	DECIDER 0, HERCULES 38, MUPET 4, OC_legacy 5	chi2 p = 2.3e-08
tx_strategy	87 PDS / 26 NACT (77% PDS)	17 PDS / 30 NACT (64% NACT)	Fisher p = 1.5e-06, OR = 5.90
survival_status	63 alive / 50 dead	4 alive / 43 dead	Fisher p = 7.4e-09, OR = 0.07
parpi	42 ND / 33 No / 38 Yes	6 ND / 30 No / 11 Yes	chi2 p = 1.3e-04

Two of these are structural rather than merely unbalanced. seq_set is nested perfectly inside platform — no sequencing set contains samples from both arms — so "platform" and "sequencing batch" are the same partition here and cannot be separated. DECIDER is 100% NovaSeq (0 of 47 non-NovaSeq samples). And the arms differ overwhelmingly in survival: 91% of non-NovaSeq samples are dead versus 44% of NovaSeq.

The treatment-strategy confound is already documented in the red-team brief (item 6) and is why the full analysis carries a treatment-adjusted secondary test. The seq_set nesting, the DECIDER asymmetry and the survival difference are recorded here because they bear on how the arm-differential gene list may be described: it is a "batch/cohort/treatment/survival-associated" list, and calling it a platform effect would overstate what this design can support.

Provenance

field	value
job	ePouta 2602738 (omentum-canary-expr), partition arkku, 8 CPU / 48 G
pipeline	Shiba v0.8.2 SIF (sha256 in the job log), patched expression.py for -s 2
alignment	STAR 2nd-pass BAMs, 202503011_ome_site_data_analysis
annotation	Ensembl 113 (homo_sapiens.gtf, 1,738,769,140 bytes)
canary draw	seed 20260918, 20 NovaSeq + 20 non-NovaSeq (all 5 non-NovaSeq platforms represented)
gates	gates.json (attached), figure expression_gates.png (attached), raw log shiba_run.log (attached)

gates.json is the machine-readable form of every gate and diagnostic above. The figure shows library sizes, TPM column sums against the 1e6 line, and the three strandedness assignment rates.