Analysis (runnable): do KMT2A partners rank by nuclear contact frequency in the same order as they rank in patients?
also called partner frequency versus contact frequency correlation · does proximity predict partner choice at KMT2A
KMT2A fuses to some partners far more often than others, and the usual explanation is that those partners sit closest to it in the nucleus. That correlation has a published clinical ranking on one side and available contact data on the other, and nobody has plotted one against the other.
The same thing, for a scientist
The clinical partner ranking is published and large: across 3401 patients, the seven most frequent translocation partner genes plus partial tandem duplication account for more than 90% of all KMT2A recombinations, 37 partners recur and 63 have been seen only once [E1]. The contact side is available from any Hi-C dataset in a haematopoietic cell. The rank correlation between them has never been computed, and this graph's `mechanism-3d-genome-proximity-partner-choice` lists it as an open question in two of its four open questions. The analysis is a Spearman correlation with a carefully built null, and its value is in the negative: if contact frequency does NOT predict partner rank in any haematopoietic cell type, then 'proximity explains partner choice' - repeated across reviews of this disease - is an analogy from other loci rather than a result at this one.
The design, and the three things that will break it1 deeper
Estimand. Spearman rho between (a) partner rank by clinical frequency in the recombinome 1 and (b) partner rank by trans contact frequency with the KMT2A locus, computed per haematopoietic cell type.
Break 1: the ranking is one centre's referral structure. The 3401-patient series comes from a single diagnostic centre, so partner frequencies inherit its referral pattern and its assay's partner coverage 1. A partner that the diagnostic panel cannot detect has a clinical frequency of zero for reasons that have nothing to do with biology. This is the same ascertainment problem that ana-repeat-feature-enrichment-kmt2a-breakpoints handles with an assay-matched null and it must be handled here too, minimally by restricting to partners the assay demonstrably detects.
Break 2: contact frequency is confounded with everything. Highly expressed, gene-dense, A-compartment loci have higher trans contact with each other by default. KMT2A partners are overwhelmingly transcription-elongation and chromatin genes, which are exactly the loci that sit in the active compartment. The null must therefore be expression- and compartment-matched, not random loci. Without that, the correlation will be positive and meaningless.
Break 3: lineage. The partner distribution differs between lymphoid and myeloid KMT2A-r disease, and the graph holds an entry on partner frequency by age and lineage. Pooling them and correlating against a single cell type's contact map mixes two distributions. Run it separately for the B-lymphoid and myeloid partner rankings against lineage-matched contact maps.
The stronger version of the test
The rank correlation is the headline but a partial correlation is the real test. Compute contact frequency, partner-locus break rate proxy (transcriptional output, TOP2 occupancy where available, DSB density from published break maps) and partner clinical frequency, then ask whether contact still predicts clinical frequency after conditioning on the break-rate proxy. That is the operational form of the contact-versus-break-frequency dispute 23 applied to this specific fusion, and it is the only version of the analysis that could actually adjudicate it.
Why a negative result here is the more valuable outcome
The proximity explanation for KMT2A partner choice is repeated widely and rests on a single FISH observation of 2-3% colocalisation in transformed lines 4. If the rank correlation is absent or weak, the graph gains a bounded, quantitative negative against a claim it currently carries at moderate confidence, and the field's attention moves to break rate and to selection on the fusion product - where the comparative evidence already points when both terms are measured 3.
If the correlation is strong, the prediction generalises immediately: partners that are contact-close but clinically rare become candidates for under-ascertainment, which is directly checkable against the 63 partners seen only once 1.
| script | analyses/kmt2a-partner-contact-correlation/run.py |
|---|---|
| status | planned |
| estimand | Spearman rho between clinical partner rank and contact-frequency rank, per cell type; plus partial correlation conditioning on a break-rate proxy |
| null model | expression- and compartment-matched control loci, not random genomic positions |
| stratification | B-lymphoid and myeloid partner rankings analysed separately against lineage-matched contact maps |
| cost tier | low |
| result summary | not run |
inputs
[ "KMT2A partner frequency ranking from the 2023 recombinome (3401 patients), split by lineage", "trans contact frequency between the KMT2A locus and each partner locus, from Hi-C in lineage-matched haematopoietic cells", "expression and A/B compartment assignment per partner locus, for the matched null", "break-rate proxies per partner locus: transcriptional output, TOP2 occupancy, published DSB density" ]
known confounders
[ "single-centre referral structure of the clinical ranking", "A-compartment co-enrichment of transcription/chromatin genes, which are what KMT2A partners are", "transcriptional output is simultaneously a contact proxy and a break-rate proxy" ]
- A strong correlation would be an explanation with no handle on it: nuclear positioning in a fetal progenitor cannot be modified therapeutically.
- A weak correlation would be reported as a null and could be a power failure instead, because trans contact estimates are noisy and seven partners is a small number of points for a rank test. The analysis must pre-specify what rho it is powered to detect.
- Do the seven commonest partners rank by contact frequency in the same order as by clinical frequency in ANY haematopoietic cell type? Never computed.
- Are any of the 63 single-occurrence partners contact-close to KMT2A? If so, they are candidates for under-ascertainment rather than for biological rarity, and that is checkable.
- Does the partner ranking differ between contact maps from lymphoid and myeloid progenitors in a way that matches the lineage split in the clinical partner distribution? That would be a much stronger result than a pooled correlation.
- Is there any haematopoietic Hi-C dataset with enough depth for trans contact estimates at single-locus resolution, or does this analysis require new data? This should be checked before the analysis is scheduled.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Across 3401 acute leukaemia patients with determined genomic breakpoints, the seven most frequent KMT2A translocation partner genes plus partial tandem duplications account for more than 90% of all recombinations, 37 partners occur recurrently and 63 have been identified only once; 107 in-frame KMT2A gene fusions are known in total.
The KMT2A recombinome of acute leukemias in 2023. · Meyer C, Larghero P, Almeida Lopes B, et al. · Leukemia 2023What this evidence cannot show
A single referral-based diagnostic centre over 20 years. Partner frequencies reflect that centre's referral pattern and its assay's partner coverage, so the ranking is not a population incidence ranking and partners the panel cannot reach are invisible.
Evidence that argues against this (1)
- An absence of detection is not an absence of events, and this graph's standing rule applies directly: a targeted assay returns negative for every partner it was not designed around. Any correlation computed against this ranking is conditioned on the assay, and a partner that is contact-close but clinically absent may be an ascertainment artefact rather than a failed prediction. Meyer C et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
With DSB numbers normalised genome-wide by ionising radiation, translocation frequency in mouse pro-B cells is directly related to pre-existing Hi-C spatial proximity, in cis and within subchromosomal domains.
Spatial organization of the mouse genome and its role in recurrent chromosomal translocations · Zhang Y, McCord RP, Ho YJ, et al. · Cell 2012What this evidence cannot show
Mouse, one cell type, and the geometry term is isolated by design through irradiation, which removes the break-rate variation that the KMT2A case is likely to be dominated by. Cited here as the statistical precedent for the analysis rather than as evidence it will succeed.
Evidence that argues against this (1)
- When contact and break frequency are measured together rather than one being normalised away, recurrent site-directed damage produces translocations proportional to break formation rate, and the authors extend this specifically to translocations driving B-cell malignancies. The contact-only correlation this analysis computes is therefore known in advance to be a partially confounded estimate, which is why the partial-correlation version is the substantive test. Hakim O et al. 2012
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Non-targeted rearrangements reflect nuclear organisation whereas break formation rate governs the location and frequency of recurrent translocations, including those driving B-cell malignancies.
DNA damage defines sites of recurrent chromosomal translocations in B lymphocytes · Hakim O, Resch W, Yamane A, et al. · Nature 2012What this evidence cannot show
Mouse B lymphocytes with engineered site-directed damage; KMT2A-rearranged leukaemia is assigned to the recurrent-damage class by analogy. The break rate at any KMT2A partner locus in a human progenitor has never been measured, which is why the break-rate term in the partial correlation must be a proxy.
Evidence that argues against this (1)
- Genes that recurrently translocate to drive leukaemias are reported as both highly transcribed AND enriched at loop anchors, which means transcriptional output is simultaneously a break-rate proxy and a contact proxy. Conditioning on it may therefore remove the very signal the analysis is trying to detect, and the partial correlation must be reported alongside the raw one rather than instead of it. Gothe HJ et al. 2019
- E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
The only direct measurement of KMT2A-partner proximity is 2-3% of transcribing KMT2A alleles lying close to AF9/MLLT3 or AF4/AFF1 by FISH in human cell lines.
Model for MLL translocations in therapy-related leukemia involving topoisomerase IIbeta-mediated DNA strand breaks and gene proximity · Cowell IG, Sondka Z, Smith K, et al. · Proceedings of the National Academy of Sciences of the United States of America 2012What this evidence cannot show
Two partners only, in transformed lines, scored as binary colocalisation of transcribing alleles. It supplies two points on a curve that needs at least seven, which is the whole reason this analysis is needed.
Evidence that argues against this (1)
- The same paper concludes that local topoisomerase II concentration does not explain KMT2A breakpoint clustering, so its authors explicitly treat proximity as one term in an incomplete model. Citing the 2-3% figure as support for a proximity-determines-partner-choice claim over-reads it, and this analysis exists because the claim has never been tested at the level of the full partner distribution. Meyer C et al. 2023
computes the rank correlation that entry lists twice as an open question
tests the hypothesis's partner-choice prediction and can falsify it
- depends on →Which partner, in which patient: KMT2A partner frequencies across six age-lineage strata
uses the lineage- and age-split partner ranking rather than a pooled one
supplies an unbiased partner spectrum to place beside contact frequency and clinical rank