Analysis: test whether published KMT2A::AFF1 occupancy is enriched at scaffold-attachment and cruciform-prone DNA beyond what CpG islands explain
also called SAR overlap permutation test · inverted repeat enrichment KMT2A · topology covariate analysis
Take the existing maps of where the fusion sits on the genome and ask whether it prefers stretches of DNA that are structurally strained — the kind that forms cross-shaped structures — over and above its known preference for unmethylated CG islands. No new experiment, only existing public data.
The same thing, for a scientist
A permutation-based enrichment test with covariate control, runnable in days on data already public. Take KMT2A::AFF1 CUT&RUN or ChIP-seq peaks from SEM/RS4;11 (and MLL-AF9 peaks from THP-1 as a second fusion), and score overlap against three annotation tracks: (a) computationally predicted cruciform-extruding inverted repeats (palindrome detection with spacer and arm-length thresholds, filtered by predicted extrusion free energy); (b) base-unpairing regions / predicted matrix-attachment regions; (c) as a positive-control feature, the annotated KMT2A breakpoint cluster region and its flanking SARs. The essential methodological point is the **covariate control**: CpG islands, GC content, chromatin accessibility and gene density all correlate with each other and with the fusion's known target set, so a naive overlap test will return significance for almost any feature. The analysis therefore compares against a matched null generated by circular-genome shifting within accessibility-and-GC-stratified bins, and additionally reports a conditional model in which topology features are added to a logistic model already containing CpG-island and ATAC covariates, so the reported effect is the **incremental** explanatory power of DNA topology. Output: odds ratio and bootstrap CI for each topology feature, the incremental AUC over the CpG-only model, and a per-chromosome sensitivity check.
The self-referential test, run as a single pre-specified hypothesis
Separately from the genome-wide scan, the analysis asks one pre-specified question with a single test and no multiple-comparison ambiguity: is wild-type KMT2A occupancy enriched at the KMT2A breakpoint cluster region itself? This is the risky prediction of hyp-kmt2a-address-is-dna-structure-not-at-richness, and it is answerable from published wild-type KMT2A ChIP-seq in any human cell type, against a null of matched-expression, matched-GC gene bodies.
It should be run and reported whatever the genome-wide result, and it should be pre-registered as a one-sided test, because a positive here is the difference between an interesting correlation and a mechanism that ties the protein's address to its own gene's fragility.
Why the covariate control is the whole analysis
This graph has already been burned by joins and inferences that looked significant and were not. Cruciform-prone inverted repeats and AT-rich base-unpairing regions are, genome-wide, anti-correlated with CpG islands — which makes the naive test conservative in one direction and treacherous in the other: a null result could be an artefact of the fusion's overwhelming CpG-island preference swamping a real secondary signal.
Hence the conditional design. The question is not 'does the fusion overlap SARs' but 'given where the fusion already is because of CXXC, does topology add anything'. Reporting an unconditional odds ratio for this comparison would be close to meaningless, and if the analysis is run by anyone else that should be the first thing checked.
A second trap: SAR annotations in the literature derive from extraction-based experiments of the 1990s and from prediction tools of variable quality. The analysis should report results for at least two independent SAR/BUR prediction methods and treat agreement between them as part of the result.
| script | analyses/sar-cruciform-overlap-kmt2a/run.py |
|---|---|
| status | planned |
| result summary | not yet run; outputs per-feature odds ratios with bootstrap CIs against an accessibility- and GC-stratified circular-shift null, the incremental AUC of topology features over a CpG-plus-accessibility logistic model, agreement between two SAR/BUR predictors, and a single pre-specified one-sided test of wild-type KMT2A enrichment at its own breakpoint cluster region |
| figure | /figures/kmt2a-topology-enrichment.svg |
| tests hypothesis | hyp-kmt2a-address-is-dna-structure-not-at-richness |
| primary risk | unconditional enrichment tests will be significant for almost any genomic feature; only the incremental, covariate-conditioned effect is interpretable |
inputs
[ "published KMT2A::AFF1 CUT&RUN/ChIP-seq peaks (SEM, RS4;11) and MLL-AF9 peaks (THP-1), hg38", "published wild-type KMT2A ChIP-seq, hg38, any human cell type, for the self-referential test", "predicted cruciform-extruding inverted repeats (palindrome scan with arm-length, spacer and extrusion-energy thresholds)", "two independent base-unpairing-region / matrix-attachment-region prediction tracks", "hg38 CpG islands, GC content in matched bins, matched ATAC-seq accessibility", "KMT2A bcr coordinates lifted to hg38 with flanking SAR annotation" ]
- A positive incremental result would still be correlative and would not establish that the AT-hooks cause the enrichment; the triple-mutant arm of the paired experiment is what converts correlation into attribution, and the analysis should be reported as a prioritisation tool rather than as evidence.
- Publishing a topology-enrichment finding without the two-predictor agreement check would repeat the kind of silent-join failure this project has already had to correct twice.
- Do the two SAR/BUR prediction methods agree well enough to support a conclusion? If they disagree substantially, the analysis's main output is that the annotation layer is not fit for purpose — which is itself worth reporting before anyone builds on SAR claims.
- Should the same test be run on the recurrent partner genes' breakpoint regions? If SAR architecture is shared across recurrent partners, that is a sequence-level model of partner choice competing with the 3D-proximity model the graph already records as failing its direct test.
- Is the fusion's occupancy at any topology feature dependent on transcription? Repeating the overlay on data from transcription-inhibited cells would separate a static DNA-architecture preference from a supercoiling state generated by transcription itself.
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
11q23 breakpoints within the MLL breakpoint cluster region correlate with scaffold attachment regions and topoisomerase II consensus binding sites, providing the annotation target for the self-referential test.
Distribution of 11q23 breakpoints within the MLL breakpoint cluster region in de novo acute leukemia and in treatment-related acute myeloid leukemia: correlation with scaffold attachment regions and topoisomerase II consensus binding sites · Broeker PL, Super HG, Thirman MJ, Pomykala H, Yonebayashi Y, Tanabe S, Zeleznik-Le N, Rowley JD · Blood 1996What this evidence cannot show
1996 SAR mapping; coordinates must be lifted to hg38 and treated as approximate, and the analysis should therefore also use modern BUR prediction rather than relying on the published SAR calls alone.
Evidence that argues against this (2)
- The correlation this analysis is built on has been reported with the OPPOSITE sign by an independent group, and that is close to fatal for the overlap statistic as designed. Hensel, Gillert, Fey and Marschalek mapped S/MARs across both the MLL and AF4 breakpoint cluster regions by halo mapping and by scaffold reassociation, then used Distamycin A competition to grade each fragment's matrix-binding AFFINITY. Comparing affinity with the known positions of patient t(4;11) breakpoints they found that DNA areas with strong MAR affinity contained FEWER translocation breakpoints while areas with weak or absent MAR affinity showed a higher breakpoint density - i.e. breakpoints in both genes are preferentially clustered OUTSIDE high-affinity matrix attachment regions. Strissel's co-localisation (E1) and this inverse correlation cannot both be simple truths about the same interval; the most likely reconciliation is that 'SAR' as scored by binary presence/absence and 'SAR' as scored by affinity are different variables, which means an overlap analysis that uses a binary SAR track is measuring an artefact of the scoring. Read: abstract only. Hensel JP et al. 2001
- The compartment the analysis's independent variable refers to may not exist. Pederson's review of the nuclear matrix traces the fraction to a 1948 patent and a 1974 relaunch and concludes that convincing evidence for its proposed role in gene expression has been elusive, that the harsh extraction and precipitation-prone conditions used to prepare it are a reasonable rather than a merely rhetorical objection, and that considerable scepticism continues to surround the nuclear matrix fraction as an accurate representation of the in-vivo situation. A scaffold/matrix attachment region is operationally defined by reassociation with that fraction. If the fraction is a precipitation artefact, a SAR track is a map of sequence properties (AT-richness, unwinding propensity) under one extraction protocol, and any overlap with fusion occupancy is most parsimoniously explained by shared sequence composition. Read: abstract only. Pederson T 2000
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
The MLL AT-hook region binds cruciform DNA and recognises structure rather than sequence, which is the premise generating the topology annotations tested here.
11q23 translocations split the 'AT-hook' cruciform DNA-binding region and the transcriptional repression domain from the activation domain of the mixed-lineage leukemia (MLL) gene · Zeleznik-Le NJ, Harden AM, Rowley JD · Proceedings of the National Academy of Sciences USA 1994What this evidence cannot show
A single old gel-shift study; the analysis tests a prediction derived from it rather than confirming it.
Evidence that argues against this (1)
- Searched and empty, recorded explicitly. The MLL AT-hook cruciform-binding result (PMID 7938000) appears never to have been independently replicated in the 30 years since, and no genome-wide map of cruciform or non-B DNA has been intersected with KMT2A or KMT2A-fusion occupancy. PubMed searches on 2026-09-18 for MLL/KMT2A with cruciform, four-way junction, non-B DNA, inverted repeat and Z-DNA found no replication and no follow-up binding study. The analysis is therefore resting its independent variable on a single 1994 gel-shift; that is not a reason not to run it, but the result should be pre-registered as a test of the 1994 claim rather than as an application of it. No citation is attached because there is nothing to cite: this note is a bounded absence plus a scope argument about a single unreplicated 1994 result, not a sourced fact. Uncited and unciteable look identical in this field, so it is said here explicitly (Q42 citation-integrity repair, 2026-09-20).
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The CXXC domain binds non-methylated CpG with an affinity around 33 nM and discriminates against methylation, establishing CpG-island content as the dominant covariate that any topology enrichment must be measured against.
The MT domain of the proto-oncoprotein MLL binds to CpG-containing DNA and discriminates against methylation · Birke M, Schreiner S, Garcia-Cuellar MP, Mahr K, Titgemeyer F, Slany RK · Nucleic Acids Research 2002What this evidence cannot show
In-vitro affinity on naked DNA; used here to justify the covariate structure of the model, not as a quantitative parameter in it.
Evidence that argues against this (1)
- Weakens the assumption that CXXC occupancy is what an overlap analysis would be predicting. Risner et al. swapped the MLL CXXC domain in MLL-AF9 for CXXC domains from DNMT1, CGBP/CFP1 and MBD1 and for the MBD1 MBD, and all of them - including domains that cannot bind the relevant DNA in vitro - still targeted the fusion to Hoxa9 in primary bone-marrow progenitors. The domain that discriminates leukemogenic from non-leukemogenic is doing something downstream of binding (protecting specific CpGs from methylation), so a CXXC affinity constant is not a predictor of where the fusion sits. Read: abstract only. Risner LE et al. 2013
- ← related toEach partner prefers a different KMT2A exon: partner-specific break position in 3,401 patients
Counter-evidence bounding the SAR / topoisomerase-II-consensus axis: etoposide-induced double-strand breaks in MLL colocalised with the RNA Pol II binding site and transcription initiation region, and explicitly NOT with a nearby topo II consensus sequence (Scharf 2007, PMID 16983345). Consensus-site overlap is a sequence annotation, not a measured cut site.
Counter-evidence (Hensel 2001, PMID 11527155): t(4;11) breakpoints in MLL and AF4 cluster preferentially OUTSIDE high-affinity matrix attachment regions — the reported SAR/breakpoint correlation is inverted when affinity rather than presence is scored
covariate-conditioned enrichment of fusion occupancy at topology features
pre-specified one-sided test of wild-type KMT2A enrichment at its own bcr