Analysis: score fetal, myeloid-priming and differentiation-stage signatures across public KMT2A-r datasets
also called lineage-switch signature scoring · fetal program score · myeloid priming score
Before running any new experiment, take the leukaemia gene-activity data that is already public, and score every case for how fetal it looks, how myeloid-primed it looks, and how immature it looks. Then check whether those scores track with which patients do badly.
The same thing, for a scientist
A computational analysis, runnable now from published data, that builds three orthogonal signature scores for KMT2A-rearranged leukaemia cases: a fetal-program score derived from the fetal-versus-childhood MLL-AF4 comparison, a myeloid-priming score derived from the shared ALL/AML regulatory core, and a differentiation-stage score reproducing the five-cluster integrative classification. The analysis asks whether any of these predicts the cases that later switch lineage or relapse, and whether the scores are independent of age and fusion partner. It is the cheapest available test of several claims in this beat and should be run before committing to the wet-lab experiments.
What is scored and from what1 deeper
1. Fetal-program score. Derived from the comparison showing that fetal-specific gene expression programs are maintained in MLL-AF4 infant-ALL but not in MLL-AF4 childhood-ALL, and that a CRISPR t(4;11) in primary human fetal liver cells drives infant-ALL-specific and fetal-specific programs 1. The score asks how fetal each case looks, continuously, rather than by age band.
2. Myeloid-priming score. Derived from the core set of KMT2A-AFF1 network nodes present in both the ALL and the AML forms of the disease 2, supplemented with the lineage regulators whose activation defines a switch.
3. Differentiation-stage score. Reproduce the five integrative clusters (IRX-type differentiated, IRX-type undifferentiated, HOXA-type MLLT1, HOXA-type MLLT3, HOXA-type AFF1) and the associated RAS pathway mutation burden, which was the prognostic variable in that study 3.
4. Fetal B-progenitor reference score. A PreProB-versus-ProB discriminator built from the human fetal B-cell hierarchy, to ask which normal fetal cell each case most resembles 4.
The questions the analysis can actually answer
- Is the fetal-program score continuous with age, or does it drop off sharply, which would argue for a hard developmental window rather than a gradient?
- Does the myeloid-priming score at diagnosis identify the cases that later switch lineage? Reported switch cases are few, so this is a hypothesis-generating rather than powered question, but it is nearly free to ask.
- Are the fetal score and the differentiation-stage score independent, or is 'undifferentiated' simply 'fetal' measured differently? If they are the same axis, several published claims collapse into one.
- Does the fetal score add prognostic information beyond age and RAS burden? If not, the fetal-program story is biologically important but clinically redundant, which is worth knowing before building a therapeutic programme on it.
Inputs and honest limits
Inputs. Published expression and methylation data accompanying the infant KMT2A-r multi-omics cohort of 84 cases 3, the infant B-ALL genome-wide cohort of 124 cases 5, the fetal-versus-childhood MLL-AF4 comparison 1, and the human fetal B-progenitor hierarchy 4. Plus, where accessible, TARGET and St. Jude Cloud paediatric ALL cohorts for external validation.
Limits, stated up front. Signature scoring across datasets generated on different platforms is vulnerable to batch effects that can manufacture apparent associations, so cross-cohort results must be shown to survive within-cohort ranking. The number of documented lineage-switch cases with matched diagnostic expression data is very small, so the switch-prediction arm is exploratory. And projecting bulk transcriptomes onto differentiation references reports the population average, which in a plastic disease is exactly the quantity that hides the subpopulation of interest.
Status is planned. No result should be quoted from this node until the script exists and has run.
| script | analyses/lineage-switch-signature-scoring/run.py |
|---|---|
| status | planned |
| result summary | Not yet run. No results should be quoted from this node. |
| statistical plan | Score all cases within cohort to avoid cross-platform batch effects; test independence of fetal and differentiation-stage scores by correlation and by variance partitioning; test prognostic value in multivariable models adjusted for age and RAS pathway mutation number; treat the lineage-switch prediction arm as exploratory given the small number of documented switch cases with matched data. |
| compute | CPU only; runnable locally or on the Explorer cluster |
inputs
[ "Infant KMT2A-r multi-omics cohort expression and methylation data (PMID 36042201)", "Infant B-ALL genome-wide cohort RNA-seq and outcome data (PMID 30679323)", "Fetal versus childhood MLL-AF4 comparison gene sets (PMID 34824279)", "Human fetal B-progenitor hierarchy single-cell reference (PMID 31383639)", "TARGET and St. Jude Cloud paediatric ALL cohorts for external validation, where accessible" ]
outputs
[ "per-case fetal-program score", "per-case myeloid-priming score", "per-case differentiation-stage assignment and RAS burden", "per-case PreProB versus ProB reference similarity" ]
tests hypotheses
[ "hyp-menin-release-lineage-destination" ]
- Cross-cohort signature scoring is highly vulnerable to batch effects, so a positive association may be a platform artefact; within-cohort ranking is the only trustworthy comparison.
- Bulk transcriptome projection reports the population average, which in a plastic disease systematically hides the small subpopulation that actually causes relapse.
- If the scores turn out to be collinear, a therapeutic programme built on 'the fetal program' as a distinct target would be chasing a re-description of immaturity.
- Are the fetal-program score and the differentiation-stage score the same axis measured two ways? If so, several independent-looking published findings are one finding.
- Does the fetal-program score add prognostic information beyond age and RAS mutation burden? If it does not, the fetal biology is real but clinically redundant.
- How many documented lineage-switch cases have matched diagnostic transcriptomes in any accessible repository? That number determines whether the switch-prediction arm is worth running at all.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Fetal-specific gene expression programs are maintained in MLL-AF4 infant-ALL but not in MLL-AF4 childhood-ALL, and a CRISPR-engineered t(4;11) in primary human fetal liver haematopoietic cells drives infant-ALL-specific and fetal-specific programs, providing the gene sets for a fetal-program score.
A human fetal liver-derived infant MLL-AF4 acute lymphoblastic leukemia model reveals a distinct fetal gene expression program · Rice S, Jackson T, Crump NT, Fordham N, Elliott N, O'Byrne S, Lara Fanego MDM, Addy D, Crabb T, Dryden C, Inglott S, Ladon D, Wright G, Bartram J, Ancliff P, Mead AJ, Halsey C, Roberts I, Milne TA, Roy A · Nature Communications 2021What this evidence cannot show
The fetal gene set is defined in a small number of biological donors and in a xenograft context, so a score built from it may be partly model-specific.
Evidence that argues against this (1)
- A fetal-program score measures state, not origin. Shared mutations across an infant KMT2A-r B-ALL and its AML relapse in the same child placed the rearrangement before haematopoietic specification, and the authors conclude that 'cell of origin cannot be inferred from the transcriptional state'. A high fetal score is therefore evidence about what the blast currently resembles, and the analysis should not be written up as evidence about where the leukaemia started. Read: abstract only. Khabirova E et al. 2022
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
The KMT2A-AFF1 gene regulatory network contains a core set of nodes present in both the ALL and the AML forms of the disease, providing the basis for a myeloid-priming score.
A KMT2A-AFF1 gene regulatory network highlights the role of core transcription factors and reveals the regulatory logic of key downstream target genes · Harman JR, Thorne R, Jamilly M, Tapia M, Crump NT, Rice S, Beveridge R, Morrissey E, de Bruijn MFTR, Roberts I, Roy A, Fulga TA, Milne TA · Genome Research 2021What this evidence cannot show
Inferred network; shared ALL/AML nodes may partly reflect shared cell-line artefacts rather than shared patient biology.
Evidence that argues against this (1)
- The gene regulatory network is a reconstruction, not a measurement. It is assembled from ChIP-seq occupancy, bulk patient RNA and CRISPR essentiality in lines, and edges inferred by combining three assays inherit the weakest one: occupancy does not imply regulation, bulk RNA averages over the clusters this beat says the disease splits into, and essentiality comes from cell lines whose lineage composition this project has shown to be confounded. "A core set of nodes shared between the ALL and AML forms" is also the expected result of reconstructing a network from genes both forms express. Building a myeloid-priming SCORE on shared nodes is circular in a specific way: nodes shared between the ALL and AML forms are by construction the ones least able to discriminate between them. Oncofusion knowledge graph (internal analysis) 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Infant KMT2A-rearranged leukaemia separates into five integrative clusters with differentiation-stage assignments, and RAS pathway mutation number predicts prognosis with the undifferentiated IRX-type cluster 100% RAS-mutated, providing both the classifier to reproduce and the prognostic covariate to adjust for.
Multi-omics analysis defines highly refractory RAS burdened immature subgroup of infant acute lymphoblastic leukemia · Isobe T, Takagi M, Sato-Otsubo A, Nishimura A, Nagae G, Yamagishi C, Tamura M, Tanaka Y, Asada S, Takeda R, Tsuchiya A, Wang X, Yoshida K, Nannya Y, Ueno H, Akazawa R, Kato I, Mikami T, Watanabe K, Sekiguchi M, Seki M, Kimura S, Hiwatari M, Kato M, Fukuda S, Tatsuno K, Tsutsumi S, Kanai A, Inaba T, Shiozawa Y, Shiraishi Y, Chiba K, Tanaka H, Kotecha RS, Cruickshank MN, Ishikawa F, Morio T, Eguchi M, Deguchi T, Kiyokawa N, Arakawa Y, Koh K, Aoki Y, Ishihara T, Tomizawa D, Miyamura T, Ishii E, Mizutani S, Wilson NK, Gottgens B, Miyano S, Kitamura T, Goyama S, Yokoyama A, Aburatani H, Ogawa S, Takita J · Nature Communications 2022What this evidence cannot show
Predominantly Japanese cohort of 84 cases; a five-cluster solution at that sample size may not replicate, which is itself one of the questions this analysis would address.
Evidence that argues against this (1)
- Inferred differentiation stage is reference-dependent, and the reference changes with ontogeny: fetal liver contains many oligopotent progenitors with intermingled myeloid, erythroid and megakaryocytic fates while adult bone marrow contains almost none. Scoring infant leukaemias against an adult-derived hierarchy and against a fetal one will not give the same cluster assignments, so the choice of reference is a free parameter that must be pre-registered. Read: abstract only. Notta F et al. 2016
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The human fetal B-cell developmental hierarchy has been mapped with single-cell transcriptomics, defining CD10-negative PreProB and CD10-positive ProB reference populations whose signatures can be used to score leukaemic cases.
Discovery of a CD10-negative B-progenitor in human fetal life identifies unique ontogeny-related developmental programs · O'Byrne S, Elliott N, Rice S, Buck G, Fordham N, Garnett C, Godfrey L, Crump NT, Wright G, Inglott S, Hua P, Psaila B, Povinelli B, Knapp DJHF, Agraz-Doblas A, Bueno C, Varela I, Bennett P, Koohy H, Watt SM, Karadimitris A, Mead AJ, Ancliff P, Vyas P, Menendez P, Milne TA, Roberts I, Roy A · Blood 2019What this evidence cannot show
Reference populations are from a limited number of gestational windows and donors, so signature robustness needs bootstrap assessment.
Evidence that argues against this (2)
- Signatures drawn from sorted fetal populations inherit those populations' heterogeneity. Clonal analysis of human lympho-myeloid progenitor gates found them functionally and transcriptionally heterogeneous, with most single cells unilineage, so a 'PreProB signature' is an average over a mixture. Scores built on such signatures will be systematically compressed and their differences between leukaemia clusters attenuated. Read: abstract only. Karamitros D et al. 2018
- Expression-based scoring may miss the layer that matters. In human fetal haematopoiesis, HSC/MPP showed extensive epigenetic but not transcriptional priming, so lineage potential was visible in chromatin accessibility and not in the transcriptome. An RNA-only signature analysis can return a null while a real difference in lineage priming exists. Read: abstract only. Ranzoni AM et al. 2021
- E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
A genome-wide infant B-ALL cohort of 124 uniformly treated cases with RNA-seq, BCR repertoire and outcome data exists and reported that reciprocal-fusion-expressing patients had significantly better 4-year event-free and overall survival, providing an independent cohort with outcome covariates for validation.
Unraveling the cellular origin and clinical prognostic markers of infant B-cell acute lymphoblastic leukemia using genome-wide analysis · Agraz-Doblas A, Bueno C, Bashford-Rogers R, Roy A, Schneider P, Bardini M, Ballerini P, Cazzaniga G, Moreno T, Revilla C, Gut M, Valsecchi MG, Roberts I, Pieters R, De Lorenzo P, Varela I, Menendez P, Stam RW · Haematologica 2019What this evidence cannot show
Data availability under controlled access may delay or prevent reuse; the reciprocal-fusion survival finding is from a subgroup of 43 patients with reciprocal-fusion data and should be treated as hypothesis-generating.
Evidence that argues against this (1)
- An outcome cohort of this kind carries immunophenotypic heterogeneity that a fusion label hides: in 133 infant ALL cases about 70% were pro-B and the rest common-ALL or pre-B, and the Ig/TCR maturation pattern differed sharply by MLL partner, with MLL-AF9 cases more mature than MLL-AF4. Differentiation-stage scores validated across a mixed infant cohort will partly recover partner identity rather than biology specific to KMT2A::AFF1. Read: abstract only. Jansen MWJC et al. 2007
- ← depends onExperiment: go back to the diagnostic sample and look for the myeloid clone before it existed
uses the existing myeloid-priming and differentiation-stage signatures to define the precursor state
the reference populations this analysis scores against are themselves clonally heterogeneous and ontogeny-dependent, so the choice of fetal versus adult reference is a free parameter that must be fixed in advance
cheapest first test; asks whether myeloid-priming score at diagnosis marks the switchers
scores the fetal program continuously and tests whether it adds information beyond age and RAS burden
- tests →Multipotent progenitor (MPP) — the stage where the undifferentiated KMT2A-r infant leukaemias map
reproduces the five-cluster differentiation-stage classification and its RAS burden association
the marker-order question any switch-signature scoring would need an answer to
blocked on the same agreement