Analysis (runnable): does product viability leave a signature in the KMT2A breakpoint map once DNA fragility is controlled for?
A computer analysis that takes every published KMT2A break position, maps it onto the protein, and asks whether breaks avoid the folded parts of the protein even after you account for the parts of the DNA that are known to be fragile.
The same thing, for a scientist
A positional regression of KMT2A breakpoint counts on genomic coordinate, with fragility covariates (intron length, Alu/repeat density, topoisomerase II consensus sites, scaffold attachment regions, GC content, DNase accessibility in relevant progenitors) and a protein-derived covariate marking whether each position encodes an inter-domain linker or a folded module. The estimand is the linker coefficient after fragility adjustment. Comparison sets are recurrent versus non-recurrent breakpoints, and leukemic versus unselected breaks where data can be obtained. All inputs are public; compute is laptop-scale.
Inputs and how to get them
- Breakpoint positions: the 2023 and 2024 KMT2A recombinome papers and their supplementary tables 1; the Frankfurt Diagnostic Center of Acute Leukemia holds the primary sequences and would need to be approached for case-level coordinates beyond what is published.
- KMT2A gene model: pinned explicitly to the 2024 38-exon consensus structure, since the 2023 and 2024 models differ 1.
- Domain boundaries: UniProt Q03164 plus Pfam, with a sensitivity analysis over AlphaFold pLDDT as a continuous, annotation-free proxy for foldedness, which avoids the circularity of hand-curated domain calls.
- Fragility covariates: RepeatMasker Alu density, published topoisomerase II consensus cleavage motifs, scaffold/matrix attachment region annotations for the KMT2A breakpoint cluster region, GC content, and DNase/ATAC signal in fetal liver and cord-blood progenitors.
- Comparators: partner-gene breakpoints (AFF1, MLLT3, MLLT1) from the same recombinome; BCR::ABL1 as a domain-poor negative control.
Statistical plan, written before looking
Poisson or negative-binomial regression of break counts in fixed-width bins across the KMT2A locus, offset by bin width. Primary model: counts ~ linker_indicator + intron_length + Alu_density + TOP2_motif_count + SAR + GC + accessibility. Report the linker coefficient with a confidence interval, not a p-value alone. Pre-register the domain model and the bin width. Run the identical model with the AlphaFold-pLDDT continuous covariate as a pre-specified sensitivity analysis. Test collinearity explicitly (variance inflation factors); if the linker indicator and intron architecture are collinear beyond a pre-set threshold, report that the question is unanswerable in this design rather than reporting a coefficient.
Secondary analyses: recurrent versus singleton breakpoints; partner-side breakpoints; the BCR::ABL1 negative control.
What would make this worth a day of someone's time
It is cheap, it is fully public, it produces a number nobody has produced, and it either strengthens or kills a hypothesis that bears on whether break position deserves to be a stratification variable. The honest expectation is a wide confidence interval crossing zero, in which case the correct output is a clear statement that the data cannot separate the two mechanisms — which is itself a useful result, because it tells the field to stop asserting either explanation confidently.
| script | analyses/breakpoint-linker-vs-fragility/run.py |
|---|---|
| status | planned |
| result summary | Not yet run. No result should be quoted from this node. |
| cost tier | low |
| timeline | 1-2 weeks of one analyst, longer if case-level coordinates must be requested from Frankfurt |
| compute | laptop-scale; no GPU |
| preregistration note | The domain model, bin width and covariate list must be fixed in writing before the regression is run, because the temptation to adjust the domain model until the coefficient is significant is the single largest threat to the validity of this analysis. |
| reproducibility note | Pin the KMT2A gene model version, the UniProt release, the AlphaFold model version and the RepeatMasker track date in the output manifest. |
inputs
[ "KMT2A recombinome breakpoint tables (PMID 37019990, PMID 38678092)", "UniProt Q03164 domain annotation and Pfam", "AlphaFold DB model for KMT2A fragments, pLDDT as a continuous foldedness covariate", "RepeatMasker / UCSC Alu and repeat tracks over the KMT2A locus (hg38)", "Published topoisomerase II consensus cleavage motifs and scaffold attachment region annotations for the KMT2A breakpoint cluster region", "Fetal liver and cord blood progenitor ATAC/DNase signal", "BCR::ABL1 breakpoint distribution as a domain-poor negative control" ]
outputs
[ "linker coefficient with confidence interval, before and after fragility adjustment", "variance inflation factors documenting whether the question is answerable in this design", "sensitivity analysis using AlphaFold pLDDT instead of curated domain calls", "recurrent-versus-singleton and partner-side comparison tables" ]
tests hypotheses
[ "hyp-severability-linker-constrained-breakpoints" ]
- A poorly specified domain model could manufacture a significant linker effect out of intron architecture, producing a confident wrong answer that then propagates into stratification arguments.
- Are base-pair-resolution breakpoint coordinates for the full 3,401-case cohort obtainable, or does the analysis have to run at subregion resolution and lose most of its power?
- Is there any published map of KMT2A breaks in cells that have not undergone leukemic selection at sufficient depth to serve as the control arm?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Published KMT2A breakpoint distributions exist at the resolution this analysis requires: 3,401 patients with subregion-level break positions and a stated 38-exon consensus gene structure in the 2024 update, in which 90.85% of breaks fall between exons 9 and 12 and 1.85% between exons 21 and 25.
The KMT2A/MLL consensus gene structure: a comprehensive update for research and diagnostic implications · Meyer C, Larghero P, Lopes BA, Marschalek R · Leukemia 2024What this evidence cannot show
Published tables are subregion-level for most cases; base-pair-resolution coordinates for the full cohort are held by the Frankfurt group and would require a data request.
Evidence that argues against this (1)
- The analysis's discriminating comparison requires a set of KMT2A breaks that never passed through leukaemic selection, and the closest published approximation already points the wrong way: etoposide applied in vitro to normal cord blood and to normal infant and child peripheral blood mononuclear cells produced MLL double-strand breaks mapping into the breakpoint cluster region (telomeric intron 8, near exon 9), with no difference in incidence between cord blood and other samples. Unselected breaks land in the same part of the gene. A comparison set of this kind is also scarce — no large, sequence-resolved catalogue of non-leukaemic KMT2A breaks appears to exist — so the analysis may be unable to construct its control arm at all, which is a feasibility problem worth stating up front. Ishii E et al. 2002
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The KMT2A recombinome reports 98% of breaks in a major cluster region (intron 7 to exon 13) with subregion densities that do not track subregion length, and identifies exon 11 as the functional cut point past which the PHD1-3 module is destroyed.
The KMT2A recombinome of acute leukemias in 2023. · Meyer C, Larghero P, Almeida Lopes B, et al. · Leukemia 2023What this evidence cannot show
Ascertainment bias toward cases referred for breakpoint characterisation.
Evidence that argues against this (1)
- The fragility covariates this analysis proposes to condition on are expected to absorb most of the signal, and the field's current review says so. Cowell and Austin summarise genome-wide demonstrations of etoposide-induced TOP2 DNA cleavage within the KMT2A breakpoint cluster region, TOP2 cleavage-site sequence preferences and DNA secondary-structure features as determinants of break-site selection. They also note that the therapy-related break hotspot and the centre of preferential cleavage are offset by more than 250 nucleotides, so a residual does remain after fragility — but the residual they describe is a sequence/structure residual, not obviously a protein-linker one, and any linker term in this regression will be competing with an unmodelled structural term rather than with nothing. The analysis should include DNA secondary-structure covariates, not only the length/repeat/TOP2-site covariates currently listed, or it risks attributing structure to linker geometry. Cowell IG et al. 2023
the fragility null is strong: TOP2 cleavage inside the BCR is demonstrated genome-wide, and etoposide produces MLL breaks in the same region in normal cord and peripheral blood cells with no leukaemic selection
DNA secondary-structure covariates must enter the fragility model: the residual left after TOP2 cleavage preference is a structure residual, not necessarily a protein-linker one
positional regression with fragility covariates and a pre-registered domain model
- related to →KMT2A / MLL1 — the gene that breaks
maps breakpoints onto UniProt Q03164 domain architecture
companion analysis: that one models breakpoint POSITION, this one models junction STRUCTURE
must run first: that analysis adjusts for fragility covariates this one asks whether to believe