Oncofusion

start / graph / ana-q140-the-sem-libraries-of-gse159608-sort-themselves-by-antibody-in-the-signal-across-two-protocols-and-two-operators

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q140-sem-menin-arm

Forty-one SEM libraries in GSE159608 sort themselves by antibody from the signal alone, across two protocols, two operators and ten months - so the deposit's own labels are confirmed by an instrument that does not share the labels' failure mode

A public dataset labels each experiment with which protein or chemical tag it measured; we ignored those labels, downloaded the raw data for 41 experiments on the same infant-leukaemia cell line, and asked the data to group itself. It grouped exactly the way the labels said it should, even matching experiments done ten months apart by different people, so the labels are right.

The same thing, for a scientist

GSE159608 contains 41 libraries from the SEM line (t(4;11) KMT2A::AFF1, infant-derived). We downloaded every deposited hg19 fragment BED, binned fragment midpoints at 5 kb, CPM-normalised, log1p-transformed and computed pairwise Pearson correlation; each library's nearest neighbour was then compared against the antibody named in its GEO `characteristics` field - a field never used to build the correlation. **28 of 41 libraries have a nearest neighbour carrying the same antibody**, and every one of the 13 exceptions is chemically explicable and was declared as an exemption before the numbers were read. Histone marks pair across *independent protocols*: H3K27me3 AutoCUT&Tag to manual CUT&Tag r=0.906, H3K4me3 r=0.901, H3K9me3 r=0.834, H3K4me1 r=0.817, H3K36me3 r=0.721 - different operators (DJ vs SH), different dates (2020-03 vs 2020-04/05). Menin pairs with menin (r=0.749), H4K16ac with H4K16ac (0.716), RNAP2-Ser5P with RNAP2-Ser5P (0.683-0.690), H3K27ac with H3K27ac. The exceptions: the four IgG libraries, whose nearest neighbour is arbitrary at r=0.12-0.22 (the four lowest values in the matrix - IgG is the joint outlier exactly as pre-registered); DOT1L, ENL and ELL, which nearest-neighbour to RNAP2-Ser5P at r=0.36-0.57 (all elongation machinery); and the KMT2A N-terminal AutoCUT&RUN libraries, which nearest-neighbour to **menin** (r=0.396, 0.634).

Go deeper
Why this was worth doing

The dispatch brief for this beat asserted that GSE159608 contradicts its own metadata so badly that no arm could be assigned from labels, and that the identities had to be recovered from signal. That premise was false (see ana-q140-the-sixteen-percent-metadata-contradiction-is-an-off-by-one-and-the-true-rate-is-zero). But the signal analysis was worth running anyway, and for a reason that survives the premise being wrong: a metadata audit and a signal audit are two instruments, and the second does not share the first's failure mode. Running it turned a metadata claim into a confirmed one. 12

Method, in enough detail to re-run

analyses/sem-arm-identity-from-signal/run.py. 41 libraries, hg19 fragment BEDs from GEO, gzip integrity verified on every file before binning (a truncated download suppresses the verdict rather than shrinking it), bin-and-delete so peak disk is one 6 MB file. 5 kb bins over chr1-22,X; fragment midpoints; CPM; log1p; Pearson. Per-library fragment counts are stored beside every correlation so that a depth-driven clustering would be visible: no library falls below 100,000 fragments (range 594,430 to 11,948,694) and the clustering is not a depth ranking - the deepest library (11.9M, ELL) and the shallowest (594k, IgG Rep1) are not neighbours, and the two H3K27me3 libraries that pair at r=0.906 differ 1.8-fold in depth. 1

The copy-number control, which had to be run and passed

SEM is aneuploid, so a correlation matrix could in principle be measuring karyotype rather than chromatin. Pre-registered control: IgG is the copy-number-only library by construction. Mean r of non-IgG libraries to IgG is 0.121; mean r among non-IgG libraries is 0.218. The structure is not copy number. 1

The exception that is a result, not an exception

The KMT2A N-terminal AutoCUT&RUN libraries (GSM4842309, GSM4842311) have menin as their nearest neighbour genome-wide. The N-terminal fragment is the one that carries the menin-binding motifs and is retained in every KMT2A fusion. Seeing the N-terminus co-distribute with menin at 5 kb resolution, with no peak calling and no target list, is the menin-interface hypothesis showing up as the top-level structure of an unsupervised analysis. See ana-q140-the-kmt2a-n-terminus-and-menin-are-each-others-nearest-neighbour-in-sem. 13

Facts and figures
scriptanalyses/sem-arm-identity-from-signal/run.py
statusrun
result summary28/41 libraries nearest-neighbour to the same antibody; cross-protocol mark pairing r=0.72-0.91; IgG is the joint outlier (r 0.12-0.22); copy-number control passed (0.121 vs 0.218).
bin size bp5000
genomehg19
cross protocol pairs
H3K36me30.721H3K36me3: 0.721H3K4me10.817H3K4me1: 0.817H3K9me30.834H3K9me3: 0.834H3K4me30.901H3K4me3: 0.901H3K27me30.906H3K27me3: 0.906
inputs
[
  "GSE159608 hg19_fragments.bed.gz, 41 SEM libraries"
]
cross protocol pairs — the numbers behind the chart
{
  "H3K27me3": 0.906,
  "H3K4me3": 0.901,
  "H3K9me3": 0.834,
  "H3K4me1": 0.817,
  "H3K36me3": 0.721
}
depth range fragments
[
  594430,
  11948694
]
How this could bite us back
  • This confirms labels; it does not validate the libraries. A correctly labelled library can still be a poor one, and two of the H3K4me3 libraries in this deposit are (see the VTP peak-share entry).
  • If a future beat uses this entry to justify trusting GEO titles in general, that is a misuse: the finding is about this deposit, checked directly, not about GEO.
Still open
  1. Does the same self-sorting hold at 500 bp bins, where promoter marks and menin should separate more sharply from broad marks?
  2. The four IgG libraries correlate with each other no better than with anything else (r 0.12-0.22). Is there any reproducible structure in CUT&Tag background in this line, and if so what defines it?
  3. Would this procedure detect a genuinely swapped pair of labels, or only a swapped antibody class? The test as run has no positive control in which a known mislabelling is recovered.
Every source behind this entry (3)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    Binning the 41 deposited SEM fragment BEDs at 5 kb and correlating them recovers the antibody named in each sample's characteristics for 28 of 41 libraries outright, with all 13 exceptions chemically explicable.

    sem-arm-identity-from-signal: run.py and results.json (this repository) · Q140-sem-menin-arm · Oncofusion knowledge graph 2026
    What this evidence cannot show

    Computed at 5 kb resolution, which is coarse for promoter-scale marks; a finer binning could change individual nearest-neighbour calls though it is unlikely to change the block structure. Pearson on log1p CPM; Spearman on the same bins gives the same partition and is stored in results.json. The analysis is ours and has not been peer-reviewed.

    Evidence that argues against this (1)
    • A nearest-neighbour test is a weak clustering criterion; it can be satisfied by a matrix with little global structure. The stronger statement is the cross-protocol pairing (r=0.72-0.91 between AutoCUT&Tag and manual CUT&Tag for the same mark), which a weakly structured matrix cannot produce. Janssens DH et al. 2021
  2. E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.dataset

    GSE159608 deposits per-sample hg19 fragment BED files and normalized-count bedGraphs for all 419 samples, processed with one stated pipeline (Bowtie2 2.4.1, --end-to-end --very-sensitive, hg19), which is what makes a cross-library comparison legitimate.

    GSE159608: Automated CUT&Tag profiling of chromatin heterogeneity in mixed-lineage leukemia (419 samples, hg19 fragment BEDs) · Janssens DH, Meers MP, Wu SJ, Babaeva E, Meshinchi S, Sarthy JF, Ahmad K, Henikoff S · NCBI Gene Expression Omnibus 2021
    What this evidence cannot show

    Read from the GEO sample records live on 2026-09-23 and from this repository's cached SOFT dump; the two agree. The processing description is the submitter's own and was not independently rerun from FASTQ.

    Evidence that argues against this (2)
    • One stated pipeline makes libraries COMPARABLE, not equivalent. It does not control for library depth, duplication or fragment-size distribution, and this beat's own companion entry found the two VTP H3K4me3 libraries hold 2.5-fold less of their reads in peaks than the rest. A clustering that sorts by antibody could equally sort by library quality where quality correlates with batch, and antibody and batch are partly confounded here.
    • The deposit is hg19 while most of this project's other chromatin work is hg38. Nothing in this entry required a liftover, but any comparison of these coordinates to the rest of the corpus does, and that step has not been taken or validated.
  3. E3established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper

    The KMT2A N-terminal fragment carries the menin-binding motifs and is retained in KMT2A fusion proteins; menin displacement by VTP-50469 removes MLL from chromatin at select genes.

    A Menin-MLL Inhibitor Induces Specific Chromatin Changes and Eradicates Disease in Models of MLL-Rearranged Leukemia · Krivtsov AV, Evans K, Gadrey JY, Eschle BK, Hatton C, Uckelmann HJ, et al., Armstrong SA · Cancer Cell 36:660-673.e11 2019
    PMID 31821784DOI
    What this evidence cannot show

    Krivtsov et al. demonstrate displacement in MLL-rearranged models; the occupancy experiments there are ChIP-seq in cell lines and PDX, not CUT&RUN in SEM, so the correspondence with this deposit's libraries is by mechanism rather than by replication.

    Evidence that argues against this (2)
    • The menin-binding motifs are in the same N-terminal fragment as the CXXC domain and the PAF1- and LEDGF-interacting regions, so N-terminal similarity to menin does not isolate the menin interface from the other contacts carried on the identical piece of protein.
    • 'Removes MLL from chromatin at select genes' is the published claim, and 'select' is doing heavy work: it is a minority of sites in the original report, so a genome-wide similarity measure is dominated by sites where the drug is documented to do nothing.
Connected to