Analysis (runnable): mine the KMT2A junction sequences for repair-pathway signature, and count the one junction class with a genetic control behind it
also called junction signature mining · microhomology length distribution KMT2A · 4-bp perfect homology junction count
Every KMT2A fusion leaves a scar where the two chromosomes were glued, and the shape of that scar says which repair crew did the gluing. Thousands of these scars have been sequenced for diagnostic purposes and nobody has ever tallied their shapes.
The same thing, for a scientist
The KMT2A recombinome project has determined genomic breakpoints in 3401 acute leukaemia patients and uses the junction sequences clinically for MRD monitoring [E1]. Those sequences carry a readable repair signature - blunt join, short microhomology, longer microhomology with deletion, or non-templated insertion - and no published study tallies that signature at cohort scale. The largest junction-structure analyses in this literature are n=14 (reciprocal t(4;11) junctions showing duplications, deletions and inversions under a few hundred nucleotides [E2]) and n=2 (therapy-related t(4;11) with two-nucleotide insertion and microhomologies [E3]). A PubMed search over KMT2A and its retired symbols for microhomology at a breakpoint or junction returns 10 records in total, and none is a cohort-scale classifier [E4]. This analysis specifies the count that is missing, and identifies the one junction class in this literature that has a genetic control attached to it: junctions with EXACTLY 4 bp of perfect homology, which are present in wild-type cells and absent in TDP2-null cells after etoposide, and which the authors note have been reported for a therapy-related MLL translocation [E5]. That class is therefore a marker of a TOP2-cut, TDP2-processed origin, and its frequency in de novo infant disease versus therapy-related disease is an aetiological readout that costs nothing but sequence access.
The estimand, stated before the method1 deeper
For each sequenced KMT2A genomic junction, classify it into mutually exclusive categories and estimate the category frequencies, stratified by aetiology (de novo infant, de novo older, therapy-related) and by partner gene:
- Blunt / no homology — direct join, no shared bases.
- Short microhomology, 1-3 bp.
- Exactly 4 bp of perfect homology — the TOP2/TDP2 cohesive-overhang class 5.
- Longer microhomology, 5+ bp, the alt-EJ/MMEJ-favouring range.
- Non-templated insertion at the join, with insert length.
- Complex — duplication, inversion or deletion at the join, as seen in all 14 cases of 2.
The primary comparison is category 3 between therapy-related and de novo infant cases. The secondary comparison is the full distribution against a matched null.
Why the 4-bp class is the one worth running first
Every other category in that list is a descriptive label whose pathway attribution is contested. Microhomology is NOT a reliable marker of alt-EJ: classical NHEJ uses short microhomology opportunistically, translocations in human cells are made by c-NHEJ 6, and Ku/LIG4-deficient B cells still generate substantial numbers of DIRECT joins. So counting microhomology lengths and inferring a pathway is exactly the inference the graph already warns against in mechanism-nhej-vs-alt-ej-at-fusion-junctions.
The 4-bp perfect-homology class is different because it has a genotype behind it. TOP2 cuts leave 4-nt 5'-cohesive overhangs; TDP2 liberates them ligation-competent; and the junction class that carries exactly 4 bp of identical sequence on both partners disappears in TDP2-null cells 5. It is a positive marker of a specific chemistry rather than a statistical tendency, and the same paper notes this junction type has been reported for therapy-related MLL translocation. If de novo infant junctions carry this class at the same rate as therapy-related ones, the claim that infant disease breaks by a TOP2-independent route (EndoG, apoptotic nuclease) is in trouble. If they do not, that claim gains its first junction-level support.
Bounded absence: what the literature actually contains
Searches run 2026-09-20, counts recorded with the exact query. The KMT2A retired-symbol trap was respected: ALL1 and HRX were included, which is how the 1998 review at PMID 9592986 surfaced at all.
| query | count |
|---|---|
(MLL OR KMT2A OR ALL1 OR HRX) AND microhomology | 10 |
(MLL OR KMT2A OR ALL1 OR HRX) AND microhomology AND (breakpoint OR junction) | 10 |
(MLL OR KMT2A) AND (junction sequences) AND (end joining) | 6 |
(MLL OR KMT2A) AND breakpoint AND (signature OR classifier OR machine learning) | 42 |
(4 bp homology OR four base pair homology) AND (MLL OR KMT2A) AND translocation | 2 |
The 42-record query is the only large one and it is dominated by prognostic-signature and gene-expression-signature papers, not junction-structure papers. Of the ten microhomology records, the on-beat ones are 2, 3, the apoptotic-translocation paper, and the ETV6::RUNX1 comparison; none exceeds a few dozen junctions. Bounded rejection, recorded so a future cycle does not re-open them: PMID 9726012 looks squarely on-beat from its retrieval (V(D)J recombinase-mediated deletion, etoposide, lymphoid line) and is about an hprt deletion, not KMT2A; PMID 9592986 is a 1998 narrative review and carries no junction tally.
The absence claim is bounded to title/abstract indexing. A junction tally buried in a supplementary table of a paper whose abstract does not mention microhomology would be invisible to these queries, which is exactly how the recombinome data themselves would be missed.
Where the sequences are, and the honest obstacle1 deeper
The Frankfurt Diagnostic Center of Acute Leukemia has determined genomic breakpoints in 3401 patients between 2003 and 2022 and states that the breakpoint sequences were used to monitor MRD 1. The coordinates and partner assignments are published; the junction sequences themselves are patient-identifying by construction (a genomic breakpoint is a clonal fingerprint - that is precisely why backtracking to a Guthrie card works 7). So this is not a download-and-run analysis. It requires either a collaboration with the Frankfurt group, or reconstruction from the smaller published sets, or mining of infant-ALL whole-genome sequencing where the reads spanning the junction are in controlled-access archives.
The realistic first version is therefore: (a) assemble every junction sequence published in figures and supplementary tables across the KMT2A breakpoint literature, which is on the order of 100-200 junctions; (b) classify them; (c) report the 4-bp class frequency with a confidence interval and by aetiology. That is a literature-curation task of perhaps two weeks and it would be the first such tally in existence.
The null model, because a short match happens by chance
Over a four-letter alphabet the expected run of identical bases at a random join is not zero, so an observed microhomology distribution must be compared to a per-junction permutation null built from the LOCAL sequence composition on both sides of each breakpoint, not from a uniform base frequency. The KMT2A BCR is repeat-rich 8, which raises the chance-match rate specifically here and would inflate any naive microhomology estimate. Report observed-minus-expected per junction, not raw microhomology length.
| script | analyses/kmt2a-junction-signature/run.py |
|---|---|
| status | planned |
| estimand | frequency of each junction class by aetiology; primary contrast is the exactly-4-bp-perfect-homology class, therapy-related versus de novo infant |
| null model | per-junction permutation over local sequence composition on both sides of the breakpoint; report observed minus expected microhomology, never raw length |
| query date | 2026-09-20 |
| obstacle | junction sequences are clonally identifying and are not deposited publicly; the 3401-patient resource is held by one diagnostic centre |
| cost tier | low |
| result summary | not run |
inputs
[ "junction sequences curated from published figures and supplementary tables across the KMT2A breakpoint literature (target 100-200 junctions)", "aetiology label per case: de novo infant, de novo older, therapy-related", "partner gene per case", "GRCh38 KMT2A and partner intronic sequence for the composition-matched permutation null", "controlled-access option: infant-ALL whole-genome sequencing reads spanning the junction" ]
junction classes
[ "blunt / no homology", "microhomology 1-3 bp", "perfect homology exactly 4 bp (TOP2/TDP2 cohesive-overhang marker)", "microhomology 5+ bp", "non-templated insertion", "complex (duplication, inversion or deletion at the join)" ]
queries — the numbers behind the chart
[
{
"q": "(MLL OR KMT2A OR ALL1 OR HRX) AND microhomology",
"count": 10
},
{
"q": "(MLL OR KMT2A OR ALL1 OR HRX) AND microhomology AND (breakpoint OR junction)",
"count": 10
},
{
"q": "(MLL OR KMT2A) AND (junction sequences) AND (end joining)",
"count": 6
},
{
"q": "(MLL OR KMT2A) AND breakpoint AND (signature OR classifier OR machine learning)",
"count": 42
},
{
"q": "(4 bp homology OR four base pair homology) AND (MLL OR KMT2A) AND translocation",
"count": 2
}
]- A junction classifier tells you how a fusion was made after the leukaemia already exists. It changes no treatment decision; its value is entirely in deciding which prevention mechanism is worth pursuing, and it can only rule mechanisms out, not in.
- If the 4-bp class turns out to be common in de novo infant junctions, the TOP2-independent models this graph currently favours for infant disease lose their main distinguishing prediction, and the prevention target list shrinks rather than grows.
- What fraction of published KMT2A genomic junctions carry exactly 4 bp of perfect homology, and does it differ between therapy-related and de novo infant cases? This is the primary question and it has never been asked.
- Do the reciprocal der(4) and der(11) junctions of the same patient carry the same junction class? If they differ, the two joins were made by different processes and the single-junction literature has been averaging over two events.
- Is junction class associated with partner gene? A partner-specific junction signature would argue that partner choice and joining chemistry are not independent.
- Would the Frankfurt group share junction structure summary statistics - microhomology length per case, no sequence - under a data-sharing agreement? Summary statistics are not identifying and would answer the primary question immediately.
- E1established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
Genomic breakpoints within KMT2A and the involved translocation partner genes were determined for 3401 acute leukaemia patients analysed between 2003 and 2022, and the breakpoint sequences were used to monitor minimal residual disease.
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, so partner and aetiology composition reflect its referral structure rather than population incidence. The paper reports breakpoint positions and partner identities; it does not report junction microhomology, and the junction sequences are not deposited in a public archive.
Evidence that argues against this (1)
- The same group's consensus gene-structure update makes the coordinate problem explicit: KMT2A exon and intron numbering has changed across reference builds and diagnostic conventions, so breakpoint positions pooled across two decades of publications are not automatically on a common coordinate system. Any junction tally assembled from the literature must re-map every position before comparing them, and a tally that does not say which build it used is uninterpretable. Meyer C et al. 2024
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Reciprocal breakpoints sequenced from 14 primary t(4;11) leukaemias all contained duplications, deletions and inversions of fewer than a few hundred nucleotides, interpreted as several DNA strand breaks on both chromosomes followed by error-prone repair, and explicitly not as the product of immune-system recombinases.
A DNA damage repair mechanism is involved in the origin of chromosomal translocations t(4;11) in primary leukemic cells · Gillert E, Leis T, Repp R, et al. · Oncogene 1999What this evidence cannot show
n=14, Sanger-era cloning; the entry does not restate this paper's content beyond what is needed to size the existing evidence base. It is the single largest published junction-structure analysis at this locus, which is the point.
Evidence that argues against this (1)
- Complex junctions are the generic signature of error-prone end joining rather than of a KMT2A-specific mechanism: junctions from nine patients with the other paradigm prenatal fusion show the same repertoire of microhomologies with small deletions and duplications, with no consistent V(D)J signal sequences or topoisomerase II consensus motifs, and were read by their authors as products of imprecise constitutive repair. A tally that finds complex junctions in KMT2A has therefore found a property of end joining, not a property of KMT2A, unless it is compared against exactly this kind of control fusion. Wiemels JL et al. 1999
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
Genomic characterisation of two therapy-related t(4;11) cases found a two-nucleotide insertion at one KMT2A::AFF1 junction, microhomologies at the junctions indicating NHEJ, and a KMT2A intron 11 breakpoint identical to two previously reported cases, supporting a preferential cleavage site under topoisomerase II inhibitor exposure.
Clustering of genomic breakpoints at the MLL locus in therapy-related acute leukemia with t(4;11)(q21;q23) · Hasan SK, Barba G, Metzler M, et al. · Genes, chromosomes & cancer 2014What this evidence cannot show
n=2. A shared breakpoint position between one new case and two literature cases is a suggestive coincidence at n=3, not a demonstrated hotspot, and the authors' own phrasing is 'strongly suggesting'. Both patients received anthracyclines as well as other agents, so 'topoisomerase II targeting' is a class attribution rather than a specific drug attribution.
Evidence that argues against this (1)
- An identical breakpoint position recurring across therapy-related cases is also observed in patients with NO epipodophyllotoxin exposure, which removes the simplest reading of a shared position as a TOP2-poison signature. This graph's `mechanism-top2-cleavage-at-kmt2a` records the same caution from the secondary-AML series that broke at the same hotspot without the drug. Broeker PL et al. 1996
- E4moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database
PubMed searches run 2026-09-20 over KMT2A and its retired symbols return at most 10 records linking this locus to junction microhomology, and none is a cohort-scale junction-signature analysis.
PubMed E-utilities counted-absence searches for KMT2A junction microhomology, run 2026-09-20; queries and querytranslation recorded in data.queries · Q32-break-machinery · NCBI E-utilities 2026What this evidence cannot show
Title/abstract indexing only. A junction tally reported in a supplementary table of a paper whose abstract does not use the word 'microhomology' is invisible to these queries. The alias set covers MLL, KMT2A, ALL1 and HRX but not every historical spelling (Htrx1, TRX1), and short aliases were not used because they inflate counts.
Evidence that argues against this (1)
- The graph's own standing warning applies to this very entry: a bounded, well-written absence claim is the kind most likely to be confidently wrong, because the symbol set and the indexing layer both silently limit it. Specifically, a large-cohort analysis framed at the level of 'structural variant signatures' rather than 'microhomology' would not appear in any of these queries, and mutational-signature analyses of infant ALL do exist. Andersson AK et al. 2015
- E5moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
Etoposide-induced KMT2A translocation junctions bearing exactly 4 bp of perfect homology are present in wild-type cells and absent in TDP2-null cells, and this junction type has been reported for a therapy-related MLL translocation in leukaemia.
TDP2 suppresses chromosomal translocations induced by DNA topoisomerase II during gene transcription · Gomez-Herreros F, Zagnoli-Vieira G, Ntai I, et al. · Nature communications 2017What this evidence cannot show
Full text read via Europe PMC fullTextXML (PMC5550487). A small number of Sanger-sequenced junctions in hTERT-RPE-1 cells; the authors write that 'further work is required to confirm this observation'. Most junctions in both genotypes lacked homology or had shorter microhomology, so the 4-bp class is a minority even in wild-type cells.
Evidence that argues against this (1)
- A 4-base chance match is not rare. Over a four-letter alphabet the probability that four specific bases coincide is 1/256 per offset, and a repeat-rich locus raises it further, so the 4-bp class cannot be read off a junction table without a composition-matched permutation null. The TDP2 genotype comparison establishes that SOME of these junctions are TDP2-made; it does not license classifying any individual junction that way. Echlin-Bell DR et al. 2003
- E6strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
In human cells chromosomal translocations are generated by canonical NHEJ; LIG4 or XRCC4 deficiency alters rather than abolishes them and shifts junctions toward longer deletions and more microhomology.
Chromosomal translocations in human cells are generated by canonical nonhomologous end-joining · Ghezraoui H, Piganeau M, Renouf B, et al. · Molecular cell 2014What this evidence cannot show
Designer-nuclease breaks at engineered loci in transformed human lines. Cited here only to justify the analysis design decision that microhomology length alone cannot assign a pathway.
Evidence that argues against this (1)
- The break chemistry decides the direction of the effect, and this paper's chemistry is not KMT2A's. TOP2-induced breaks carry 4-nt cohesive overhangs that religate without processing, and for those the same end-joining machinery is protective rather than translocation-generating; the authors of the TDP2 study state the contrast with nuclease-induced breaks explicitly. Any junction classifier trained on nuclease-cut data will therefore mis-assign TOP2-made junctions. Gomez-Herreros F et al. 2017
- E7established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.paper
A patient-specific genomic KMT2A::AFF1 breakpoint junction is clonally unique enough to be used as a fingerprint, and was recovered from archived neonatal blood spots of children later diagnosed with infant leukaemia.
Backtracking leukemia to birth: identification of clonotypic gene fusion sequences in neonatal blood spots · Gale KB, Ford AM, Repp R, et al. · Proceedings of the National Academy of Sciences of the United States of America 1997What this evidence cannot show
Cited here for one narrow purpose: to establish that a junction sequence is personally identifying, which is the governance obstacle to pooling the recombinome junction sequences for this analysis. It is not cited for anything about latency or in-utero origin, which the graph covers elsewhere.
Evidence that argues against this (1)
- Clonal uniqueness is not absolute: the therapy-related literature reports a KMT2A intron 11 breakpoint position identical across three independent patients. Positions can recur even where the full junction sequence does not, so 'identifying' applies to the sequence including its microhomology and insertions, not to the coordinate alone - a distinction that matters for what can be shared. Hasan SK et al. 2014
- E8strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The KMT2A breakpoint cluster region is repeat-rich: full resequencing of the 8.3 kb BCR in 87 individuals found four polymorphic sites and every one of them lay inside an Alu repetitive element.
Polymorphisms in the MLL breakpoint cluster region (BCR) · Echlin-Bell DR, Smith LL, Li L, et al. · Human genetics 2003What this evidence cannot show
An indirect argument for repeat density: four polymorphisms all falling in Alu is strong but is not a measured repeat fraction. The same study is a negative on risk - the three poly-A polymorphisms occurred at equal frequency in patients and controls.
Evidence that argues against this (1)
- Repeat content at this locus has never been compared to a matched genomic background, so 'repeat-rich' is a qualitative impression carried forward from pre-RepeatMasker-era work. The absence of that comparison is the subject of the companion analysis and applies equally to its use here as a chance-match inflator. Q32-break-machinery 2026
the cohort-scale junction classification that entry's open question asks for
counts the 4-bp perfect-homology class whose TDP2 dependence that entry records
junction structure at the KMT2A major BCR, stratified by aetiology
- related to →Do different breaks give different biology? A verdict on partner identity vs breakpoint position
a junction-class difference between de novo and therapy-related disease would be a second, orthogonal axis of 'different breaks'
companion analysis: that one models breakpoint POSITION, this one models junction STRUCTURE
inverse PCR is the only readout in this system that returns junction sequence, and flavonoid junctions were mostly MMEJ
would generate the first purpose-built junction library for the signature classification
53BP1 governs the end-protection/resection choice that sets the junction class that analysis would count