Oncofusion

start / graph / ana-ng2-in-kmt2a-rearranged-patients-is-bimodal-and-the-552-fold-median-ratio-is-an-artefact

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q138-kmt2a-aff1-surfaceome

NG2 in KMT2A-rearranged patients replicates in two independent TARGET profiles - 3 of 4 and 9 of 13 above the 90th percentile of everyone else - but it is bimodal, and the 552-fold median ratio I first computed is an artefact of a near-zero denominator

also called NG2 patient replication · the 552-fold artefact

In children's leukaemia samples, the NG2 protein's gene is switched on far higher in KMT2A-rearranged patients than in any other kind - but only in about half to three-quarters of them, not all. A first pass produced a headline 552-fold difference, which turned out to be an accident of dividing by a number close to zero, and is not reported as a result.

The same thing, for a scientist

The cell-line finding for CSPG4/NG2 replicates in patient material, in **two independent expression profiles of the same TARGET Pediatric ALL Phase II cohort**, split on the study's own `MLL_STATUS` attribute [E1]: | profile | MLL+ n | MLL- n | MLL+ values | MLL- median | MLL- 90th pct | MLL+ above MLL- p90 | exact binomial p vs a 10% null | |---|---:|---:|---|---:|---:|---:|---:| | RNA-seq RPKM | 4 | 168 | 0.01, 3.69, 39.65, 49.17 | 0.039 | 0.44 | **3 of 4** (Wilson 0.30-0.95) | 0.0037 | | microarray | 13 | 231 | 20.95 to 1754.21 | 60.53 | 141.13 | **9 of 13** (Wilson 0.42-0.87) | 4.9e-7 | **And it is bimodal.** In the RNA-seq arm one of four MLL-positive samples is at **0.01 RPKM** - flat zero - while two are above the maximum of all 168 MLL-negative samples. In the microarray arm the two lowest MLL-positive values (20.95, 44.38) are below the MLL-negative median. NG2 is not a uniform property of KMT2A-rearranged disease; it is very high in most and absent in some [E1]. **The number I nearly reported.** The first pass computed a median ratio of MLL-positive over MLL-negative and got **552-fold**. That figure is the median of four values divided by a median of 0.039 RPKM. It is arithmetically correct and it is not a result: it is a property of a near-zero denominator and an n of four. **It is recorded here so that it cannot be quoted later from a summary**, and the statistic actually used is the proportion above the negative cohort's 90th percentile, which has a null (10% by construction) and an interval.

Go deeper
Why the null had to be written down

'3 of 4 samples are high' is not a finding until you say how many you would expect. By construction, 10% of any draw falls above a cohort's own 90th percentile. Against that null, 3 of 4 gives exact binomial p = 0.0037 and 9 of 13 gives p = 4.9e-7 1. Both survive, and the microarray arm - which has three times the sample size - survives overwhelmingly.

The Wilson intervals are wide (0.30-0.95 and 0.42-0.87) because the sample sizes are small, so the honest reading is 'most, but demonstrably not all' rather than any particular percentage. That is consistent with the diagnostic literature, where NG2 positivity by flow is high but not universal in KMT2A-rearranged disease, and where its accuracy varies by centre 2.

What the two profiles agree and disagree about

They agree completely on NG2. They are also the control for each other: the microarray arm has 13 MLL-positive samples against the RNA-seq arm's 4, and the two platforms are not on comparable scales, so only rank-based statements can be carried between them - which is why the comparison above is a proportion above a percentile rather than a fold-change.

The same two profiles disagree about CD19, and that disagreement is written up separately rather than folded in here.

What this cannot support

MLL_STATUS does not name the partner. Four and thirteen MLL-positive samples cannot be subdivided by fusion partner, so this is a KMT2A-rearranged result and not a KMT2A::AFF1 result - even though the cell-line arm that motivated it is KMT2A::AFF1-specific and the published mechanism is MLL-AF4-specific 3.

TARGET ALL Phase II is high-risk childhood B-ALL, not infant ALL, so the population with the highest KMT2A::AFF1 burden is the one least represented. Expression is not surface presentation. Every expression number in this entry is mRNA. It says nothing about how much protein reaches the membrane, whether the epitope is accessible to an antibody, or whether the protein is shed. A cell line is not a patient, and the Human Protein Atlas consensus tissue panel is adult tissue while the population of concern is infants.

Facts and figures
scriptanalyses/kmt2a-aff1-surface-target-screen/run.py
statusrun
result summaryNG2 replicates in patients in two profiles and is bimodal; the reportable statistic is the proportion above the MLL-negative 90th percentile, not a fold-change.
rna seq
{
  "profile": "all_phase2_target_2018_pub_rna_seq_mrna",
  "mll_pos_values": [
    0.01,
    3.69,
    39.65,
    49.17
  ],
  "mll_neg_n": 168,
  "mll_neg_median": 0.039,
  "mll_neg_p90": 0.44,
  "mll_neg_max": 13.52,
  "above_p90": "3/4",
  "wilson_95ci": [
    0.301,
    0.954
  ],
  "exact_binomial_p_vs_10pct_null": 0.0037
}
microarray
{
  "profile": "all_phase2_target_2018_pub_mrna",
  "mll_pos_n": 13,
  "mll_pos_median": 334.47,
  "mll_neg_n": 231,
  "mll_neg_median": 60.525,
  "mll_neg_p90": 141.13,
  "above_p90": "9/13",
  "wilson_95ci": [
    0.424,
    0.873
  ],
  "exact_binomial_p_vs_10pct_null": 4.91e-7
}
withdrawn statistic
{
  "value": "552-fold median ratio (RNA-seq)",
  "why": "median of n=4 over a median of 0.039 RPKM; a near-zero-denominator artefact, not a result. Recorded so it cannot be quoted from a summary."
}
How this could bite us back
  • An antigen present in most but not all patients cannot be used without a companion diagnostic, and a companion diagnostic whose accuracy varies by centre is a trial-design problem before it is a biology problem.
  • The one MLL-positive sample at 0.01 RPKM is a patient who would be enrolled by genotype and would not respond. If the NG2-negative fraction is enriched for any particular biology, an NG2-directed therapy selects for it.
  • This entry's own headline number was wrong by two orders of magnitude on the first pass. Any reader taking the 3-of-4 figure forward should note it has a 95% interval from 0.30 to 0.95.
Still open
  1. What distinguishes the NG2-negative KMT2A-rearranged patients from the NG2-high ones? If it is the fusion partner, that is a stratification variable for any NG2-directed trial; if it is blast maturation, it is a moving target under therapy.
  2. Do the NG2-low patients become NG2-high at relapse, or the reverse? Paired diagnosis-relapse material would answer it and this graph holds no such measurement.
  3. Is the bimodality present at protein level by flow, or is it an mRNA phenomenon? The diagnostic literature reports NG2 positivity rates that would let this be checked against an existing series.
Every source behind this entry (3)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    In TARGET Pediatric ALL Phase II, CSPG4 exceeds the MLL-negative 90th percentile in 3 of 4 MLL-positive samples in the RNA-seq RPKM profile (MLL-negative n=168, median 0.039, p90 0.44) and in 9 of 13 in the microarray profile (MLL-negative n=231, median 60.53, p90 141.13); exact binomial p against a 10% null is 0.0037 and 4.9e-7 respectively; the distribution is bimodal, with one MLL-positive RNA-seq sample at 0.01 RPKM.

    What this evidence cannot show

    n = 4 and n = 13. The two profiles overlap in patients and so are not fully independent samples, only independent measurements. MLL_STATUS does not name the partner. Values are bulk tumour expression, not per-blast.

    Evidence that argues against this (1)
    • The first-pass median ratio from the same data was 552-fold, which is not a defensible statistic. Two summaries of the same measurement can differ by two orders of magnitude depending on which one is chosen, and the one that reads most impressively is the one that should be distrusted. Oncofusion Knowledge Graph 2026
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    NG2's predictive accuracy for MLL rearrangement varies greatly between centres, and proper staff training and standardised sampling are essential when receiving samples from multiple centres.

    PMID 21444110DOI
    What this evidence cannot show

    148 patients under 24 months from multiple Brazilian centres, flow cytometry. This is about a protein assay, while the measurement above is mRNA, so the two sources of heterogeneity are different.

    Evidence that argues against this (1)
    • If NG2 flow positivity varies by centre for technical reasons, an mRNA bimodality could still be a genuine biological split, and the two observations may have nothing to do with each other. cBioPortal for Cancer Genomics 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    NG2 is an epigenetically regulated direct target gene of MLL-AF4 specifically, and NG2 expression is a hallmark of MLL-rearranged B-ALL.

    PMID 39093982DOI
    What this evidence cannot show

    Abstract read in full. The mechanism is AFF1-specific; the patient measurement above is partner-agnostic, so the two cannot be joined without an assumption.

    Evidence that argues against this (1)
    • This beat's own cell-line measurement puts NG2 at only 3.05x the KMT2A-other-partner arm, so an AFF1-specific mechanism does not produce an AFF1-specific expression level, and NG2 is substantially a pan-KMT2A-rearranged antigen. Oncofusion Knowledge Graph 2026
Connected to