Repository navigation
3'UTR extension is not needed on WS298: 0.15% of reads, measured on 46.3B reads across 13 worm datasets #120
lhqing
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Closing #52 as a negative result, recorded here because the measurement is reusable and the method has two traps worth remembering.
Measured against our own corpus — 138 samples / 13 datasets / 46.3 billion uniquely-mapped
reads, every nuclear chromosome, no subsampling. A perfect, per-gene, data-driven 3′UTR extension
would recover 0.15% of unique reads at +500 bp. That does not justify a second annotation, a
STAR index rebuild, a second run of the corpus, or the operonic bleed-through risk.
The proposed benchmark was therefore not run: its premise failed. Full method, scripts and
per-sample JSON in
260612-worm/GTF-UTR3/(masks.py,measure.py,aggregate.py,results.tsv).1. Why WS298 is not the annotation the published recipes were fixing
The two worm atlases that motivated this — Ghaddar 2023 (WS260) and Purice 2025 (WS280, "too short
for reliable gene assignment") — were extending a materially poorer annotation. WS298 is not that:
Median annotated 3′UTR (153 nt) already exceeds the empirical worm median (~122 nt, UTRome v2,
10.1101/gr.254839.119). There is little systematic under-annotation left to recover. And because
this GTF is Ensembl-style — terminal exons already span the UTRs — STARsolo
Genecounts annotated3′UTRs today, so the geometry caveat raised in this issue is confirmed as the right concern and is
already satisfied by the stock annotation.
A blanket +500 bp would append 500 bp to genes whose real 3′UTR is annotated at a median of 153 nt;
only 12% of annotated 3′UTRs reach 500 nt. It would also overrun the neighbour for 18% of genes
(only 82% have ≥500 bp to the next same-strand gene; 96% have ≥100 bp) — the operon risk is real.
2. There is a 3′ read-through signal. It is directional, and it dies by ~300 bp.
Read density past an annotated end, relative to far-intergenic (>5 kb), unique primary reads only.
The 5′ control is the same measurement taken past 5′ ends over matched genomic space (1.90 vs
1.91 Mb in the first bin) — it isolates true directional read-through from generic
annotation-boundary noise.
The control sits at ~1.0 throughout, so the 3′ signal is genuinely 3′-specific — not mismapping or
fuzzy gene bounds. But it is short-range: 3.7× in the first 100 bp, ~1.3× by 300 bp,
indistinguishable from background past 500 bp. Nothing in this data supports a 500 bp extension;
the evidence only ever supported ~100–200 bp. Per-dataset breakdown in
results.md.3. What extension could actually recover: 0.15% of reads
Directional excess = (reads past 3′ ends) − (reads past 5′ ends), matched space, as a percentage of
uniquely-mapped reads. This is the ceiling: the most a perfect extension could rescue.
Best case in the whole corpus is GSE208154 at 0.35%.
Two results worth recording:
WS260. On WS298 its own directional excess is 0.20%. The extension was solving a WS260
problem, not a worm problem.
largest unassigned fraction in the corpus (19.4% of reads unique-mapped but assigned to no
GeneFullfeature) yet the lowest 3′ excess (0.02%), with 3′ enrichment below 1.0 (0.59×)and a 5′ control above it. Its unassigned reads are not 3′ read-through at all. This is why
"unassigned reads" is not evidence for UTR extension — a distinction the issue's original framing
did not draw, and the single most useful thing this exercise produced.
For context, the unassigned ceiling itself (from the saved
qc.json.gz, 138 live samples): median5.0% of reads are unique-mapped but assigned to no
GeneFullfeature (range 3.5–19.4%). Only afew percent of that is 3′-directional — the rest is antisense (we run
--soloStrand Forward),unannotated ncRNA, and intergenic/ambient signal, none of which an extension touches.
4. Method, and two bugs that would have produced the opposite answer
Both are worth stating, because either alone yields a confidently wrong result:
/usr/bin/samtoolsis 1.13and decodes only the first few containers before
[E::cram_next_slice] Failure to decode slice.Piped with stderr suppressed, that looks like a complete stream — measured, it delivered
0.13%–8.1% of chrI, biased to the chromosome start.
samtools quickcheckpasses. All readshere go through the
align-rnacontainer's samtools, with exit status checked.+read but the 3′ end of a
−read. Anchoring both on it inflated the minus-strand1-100bin~4× purely by geometry. Reads are anchored at the edge nearest the reference end, and a read is
counted as assigned if either end falls in a gene (STARsolo assigns on overlap, so a read
straddling a boundary is already counted).
Also: MtDNA is excluded (all 36 mito genes are on
+, so antisense mito reads — ~7% of uniquereads — would land in bins with a 13.8 kb denominator), and superseded runs under
seqforge/archive/pipeline/are excluded in favour of the live ones.5. The infra question is answered: no
CB/UBin the saved CRAMsrule starsolo_countruns STAR without--outSAMattributes, so the defaultStandard(
NH HI AS nM) applies. Verified across all 175 CRAMs (138 live + 37 archived): noCB,UB,CR,CY,UR,UY, and read names are plain SRA names. So saved CRAMs support pileup work likethis analysis, but cannot be re-quantified per-cell — any future annotation change means
re-alignment from FASTQ, not a cheap re-count. Not filed separately per our issue-consolidation
preference; the writeup is in
GTF-UTR3/cbub-issue.mdif it's worth folding into #56.6. What is preserved if this is ever revisited
The factorial table's entire stock-WS298 column already exists for all 138 samples in the saved
qc.json.gz—Median Gene per Cell,Median UMI per Cell,Fraction of Unique Reads in Cells,Reads Mapped to Gene/GeneFull, for all five soloFeatures. A future comparison costs only thesecond column: one STARsolo pass per sample (one alignment emits all five features), not four.
What would reopen this: a corpus on a materially worse annotation; a non-polyA or non-3′
chemistry; or a WormBase release that regresses 3′UTR calls. Not a new tool — the limit here is the
signal, not the caller, and neither peaks2utr nor GeneExt can recover reads that are not there.
Scope note: GSE266161 (Moro 2025 RoCK&ROI) was out of scope as stated. GSE274290 is BD Rhapsody WTA,
still 3′-biased, and is included (strongest enrichment at 16.2×, still only 0.26% excess).
Part of #42.
All reactions