What the multi-mapped reads actually are: the chrI rDNA array, two copies counted three ways

[lhqing]

Answering #427 on the full 784-worm plate of #425 / discussion #450. The full write-up with every
table is the dated measurement in docs/research/multimapper-composition-aging-plate.md (PR #473);
this post is the standalone record, so the run log carries the answer without needing the repo open.

The question it started from: the 16-worm pilot put a multimapper share on every cell and it ran
6.1% to 40.5% across one strain, one protocol, one sequencing run. Nothing said what those reads
were, and nothing graded the number.

The answer in one paragraph

They are ribosomal RNA. Three things compose to make the number what it is. The assembly
collapses the chrI 45S rDNA array into ~2.1 near-identical copies at the chromosome terminus, so an
rRNA read can never map uniquely. The recipe bounds no mate gap, so those two copies yield three
legal pairings and the reads come back at NH=3 rather than NH=2. And library prep sets how
many such reads a library carries, which is what varies ~8x across the plate. No one of the three is
the answer alone.

On the whole plate the STAR read-level share runs 6.06% to 51.47%, median 30.88%.

1. What they are

Three independent readings, three different denominators, one answer:

reading	denominator	result
obsm/multimapping_hits, 784 cells	all fragments (3,124,390,004 multiply placed)	NH=3 is 81.31%, NH=2 is 15.53%
multiplaced CRAM windows	all records	6.38M of ~6.5M chrI NH=3 records in day17_N2_12 at 15.06–15.07 Mb
umi_multimapping_placement	UMI-deduped, one-gene-body (36,732,376 molecules)	rRNA genes are 49.53% of placement

The locus is the chrI terminus (chrI is 15,072,434 bp):

gene	span on I__ce11	biotype
rrn-3.56	15,060,299–15,061,150	pseudogene
rrn-1.1	15,062,083–15,063,836	rRNA (18S)
rrn-2.1	15,064,301–15,064,453	rRNA (5.8S)
rrn-3.1	15,064,838–15,068,346	rRNA (26S)
rrn-1.2	15,069,280–15,071,033	rRNA (18S)

Top placed genes plate-wide, as % of the 36.7M-molecule placement total:

#	gene	%		#	gene	%
1	rrn-1.2	28.34		6	pud-1.2	2.76
2	rrn-1.1	15.57		7	pud-2.2	2.54
3	rrn-3.1	5.57		8	vit-4	2.11
4	F23A7.4	3.99		9	F23A7.8	1.80
5	rrn-3.56	3.60		10	Y45F10C.2	1.45

Ranks 1, 2, 3 and 5 are one repeat unit — together 53.09% of all placement. The tail is the
expected tandem/paralogous families: vitellogenins, histones, small heat-shock, nspc, pud, and
the F23A7.4/F23A7.8 pair 560 bp apart on chrX. Top 30 is 80.57%.

Only 25 distinct genes ever occupy any cell's top 5, out of 46,926. rrn-1.1 and rrn-1.2 are
in the top 5 of all 784. Below rank 3 the ordering does move — median per-cell profile correlates
with the plate at rho 0.71 — so the head is invariant and the tail is not.

2. Why NH=3, which is the part worth remembering

NH=3 is not three copies of anything. It is two copies counted three ways.

The array is ~2.1 tandem copies of a 7,197 bp unit, and the period reads straight off the GTF:
rrn-1.2 start − rrn-1.1 start = 15,069,280 − 15,062,083 = 7,197. Carry it back and rrn-3.56 +
7,197 lands inside rrn-3.1, so rrn-3.56 is the previous unit's 26S tail, not a gene family of
its own — its pseudogene label is an artifact of annotating a collapsed array.

A single mate off that unit has two placements. But STAR places a pair, and this recipe
sets no alignMatesGapMax, so STAR derives 589,824 bp — eighty times the repeat period. With copies
7,197 bp apart, three pairings score identically: both mates in copy A, both in copy B, and mate 1
in A with mate 2 in B. The reciprocal is RF-oriented and not emitted. Exactly three — k(k+1)/2 at
k=2.

Two experiments settle it. Same reads, single-end vs paired-end:

run	NH=2 %	NH=3 %
day17_N2_12 paired-end, as the plate ran it	14.10	83.45
day17_N2_12 single-end, mate 1, same reads	92.51	3.73

And a mate-gap sweep, which steps on the repeat period to the base pair:

alignMatesGapMax	500	7,196	7,197	7,198	20,000
NH=3 %	0.55	14.41	84.28	84.28	85.61

99.94% of NH=3 fragments at the locus carry exactly one cross-copy placement against two same-copy
ones, and |TLEN| on the cross-copy one spikes at 7,200 — one repeat period plus the insert.

Ruled out with numbers: read length (single-end tiles at 150/100/75 bp give 4.78/7.62/8.96%, and
raw reads are uniformly 150 bp anyway), a third diverged copy (95.6% of all three placements lie
inside I:15,059,000–15,072,434), and soft-clipping (an adapter-free synthetic set reproduces it
at 85.61%).

This is a diagnostic curiosity, not a defect. The cross-copy pairing only arises where each mate
is already multiply placed, so it can never turn a uniquely-placed fragment into a multi-mapped one
and never reaches the primary matrices. It is recorded so nobody reads "81% hit three loci" as
evidence the array holds three copies. It holds two.

3. Why one library has 8x another's share

Nothing you would expect. Spearman rho against the STAR share, n=784:

ruled out	rho	p
depth (fragments)	−0.029	0.41
read length	+0.003	0.93
age (continuous)	−0.035	0.32
genes detected	−0.119	rho² = 1.4%
replicate index	−0.094	negligible
bacterial share	−0.147	significant but negative — bacteria-heavy cells have fewer

The saturation hypothesis from #427's second comment is eliminated, and backwards. It proposed
the share tracks library complexity, on the lead that day17_N2_7 is both lowest-multimapper and
most-saturated. Across 784 cells rho = +0.354 — more saturated cells have more multimappers.
day17_N2_7 is an outlier against the plate, not an instance of it.

rRNA is the largest identified driver, and it is not the explanation:

definition	median	max	rho vs STAR share
rRNA share of placement	51.35%	94.37%	+0.471
rRNA share of umi_combined	3.15%	32.54%	+0.448

The second row carries the argument — there rRNA is uniquely mapped and sitting in the expression
matrices, so the correlation is not a circular consequence of rRNA dominating the placement layer.
But rho² ≈ 22%. Largest single driver; the rest is unaccounted for.

Age and strain are significant as factors but non-monotonic (day7 highest at 33.62, day13
lowest at 24.82, and day13 also holds the plate maximum) — read as a plate/batch effect with wide
within-day spread, not biology.

4. The 6.1% vs 40.5% contrast: quantity, not composition

	day17_N2_7	day17_N2_12
STAR share	6.06%	40.52%
multiply-placed fragments	503,687	3,915,330 (7.77x)
saturation	0.810	0.519
% of multiplaced at NH=3	43.09	84.05
rRNA share of umi_combined	1.09%	3.08%

Cosine similarity of the two composition vectors is 0.973; ranks 1 and 2 identical; 90% of _7's
placement sits in genes also placed in _12. Same locus classes, 7.8x the quantity.

The general form: by STAR-share tercile, high-multimapper cells have a lower mean NH (2.939 →
2.920), because the extra multimappers pile into NH=3 while every bin at NH≥4 shrinks. A cell with
more multimappers is not hitting more loci per read — it is hitting the same small-copy families more
often.

5. A trap worth carrying forward — two statistics share the name

Same shape as the bacterial-share trap this discussion already recorded.

statistic	denominator	median	max
STAR read-level (Log.final.out)	all STAR input reads	30.88%	51.47%
counter fragment-level (obs/multimapping/obs/n_fragments)	ce11 fragments only	34.87%	69.24%

rho between them is only 0.79 — they do not rank the plate the same way. The counter share
exceeds the STAR share in 784 of 784 cells, by construction: the denominators differ. Plate-wide
the counter saw 83.3% of STAR's input; per cell that runs 1.35% to 97.2%, and 33 cells are under
half.

The worked case: day17_DA_1 reads STAR 7.90% against counter 47.88%, because it is a near-pure
bacterial library — 656,460 ce11 records against 28,273,805 ecHT115 records, and n_fragments
315,143 against 23,276,339 STAR input reads. Its 47.88% is a share of almost nothing.

Never quote the counter-share maximum (69.24%) bare. That cell's counter saw 18.9% of its reads.

6. What it decided

Neither a threshold nor a report panel for multimapper share. The share is substantially a proxy
for rRNA carry-over, and grading a proxy is worse than grading the thing it proxies; the number is
composite (assembly artifact × prep variable, plus a paralogue tail) so a bar would not survive a
change of assembly or annotation; and it correlates weakly with anything actionable. So
map/star-umi and map/star-umi-chimera stay in MODULES_WITHOUT_CROSS_CHECKS — with the
argument now resting on evidence rather than on absence of evidence.

--outSAMmultNmax was not changed, in-tree or out. #427's step 3 was judged not demanded.

Proposed instead: a direct per-cell rRNA-fraction metric — which turned out to span two repos.
seqforge must not parse gene biotype itself (R10), and the derivation is genuinely not portable:
gene_biotype in WormBase/RefSeq vs gene_type in GENCODE, GENCODE splitting rRNA /
rRNA_pseudogene / Mt_rRNA where others do not, and sacCer3 carrying no biotype at all.

• An annotation can name the genes in a category, and rRNA is the first caller liulab-genome#111 — a registered annotation can name the genes in a category
• A cell can say what fraction of it was ribosomal RNA #471 — the per-cell rRNA fraction, blocked by the above
• Ribosomal RNA is counted as expression, and nothing on the object says so #472 — rRNA is counted as expression and nothing on the object says so

7. Caveats

• Chimeric arm only. Cannot separate any contribution of the chimeric reference itself.
  Cross-Component attribution is Run the in-house SMART-seq3 plate against the built chimera, and get the bacterial fraction #425's item 2 and stays separate.
• The representative-locus caveat, on every ranking. --outSAMmultNmax 1 emits one record per
  fragment, so a count in gene X means one of the loci it could have come from is X. Concretely:
  rrn-1.2 at 28.3% against rrn-1.1 at 15.6% is STAR's pick bias between two near-identical
  copies, not a 2:1 abundance difference.
• The plate is one workflow stamp behind — it ran under 2026.8.18; ADR-0056's intron cap (The intron cap reaches STAR's command line #464)
  landed after, under 2026.8.19, and that bump also began bounding the mate gap. At 50,000 bp against
  a 7,197 bp period the pairing in §2 survives, but that is reasoned from the flag's value, not
  measured under it.
• The placement layer's denominator is small — 36.7M molecules, 1.18% of the NH denominator.
• All 784 cells reported, no exclusions. Ten are under 1M fragments; restricting to the other 774
  moves exactly one number, the maximum (51.47% → 44.94%). Those ten are all rRNA-dominated
  (65–94% of placement), and day13_CF_28 — the plate maximum — is 94.4% rRNA. The cleanest single
  illustration of the whole result.

Artifacts

aging_SS3/script/metrics427/plate/ for the plate statistics and .../nh3/ for the pairing probe,
each with a README.md. Both have the scripts that produced them.