Oncofusion

start / graph / ana-q139-the-sibling-beats-mll1-verdict-survives-and-strengthens-under-the-stricter-measure

Analysisstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q139-fusion-degron

Re-scored under the measure that passes a degron positive control, cycle 117's MLL1 two-by-two verdict does not merely survive - it strengthens from a 3.33-fold to a 4.64-fold separation between the menin inhibitor and the DOT1L inhibitor, because the MOLM13 arms do not show the peak-broadening that broke the metric elsewhere

also called cycle 117 survives the metric change · 3.33 becomes 4.64 · the MOLM13 recheck

A finding published here last cycle was measured with a yardstick we have just shown can be badly wrong, so we re-measured it with the corrected one. It held up, and the gap between the two drugs got larger rather than smaller.

The same thing, for a scientist

**Why this had to be run rather than assumed.** `ana-q139-the-one-kb-bin-overlap-retention-measure-fails-its-own-positive-control` shows the 1 kb bin-overlap measure failing a degron control in GSE173599. Cycle 117's headline — the MOLM13 MLL1 2x2 — is computed with that same measure on that same deposit. **A defect found in a measure obliges a recheck of everything built on it, not an assumption that the conclusion is safe.** **Result, MOLM13 MLL1, top decile of the DMSO peak set:** | arm | bin overlap (cycle 117) | summit containment | |---|---:|---:| | EPZ-5676 1 µM | 82.5% retained | **88.3%** | | VTP-50469 100 nM | 41.7% retained | **45.7%** | | both | 15.4% retained | **17.9%** | | | bin | summit | |---|---:|---:| | DOT1L-inhibitor loss | 17.5% | 11.7% | | menin-inhibitor loss | 58.3% | 54.3% | | **ratio** | **3.33** | **4.64** | **The verdict holds and sharpens.** The two measures agree to within ~6 points on every arm, and the menin/DOT1L separation *increases* because the stricter measure takes proportionally more away from the smaller (DOT1L) effect. **And this explains why the defect was not caught before.** The MOLM13 arms **do not broaden**: their treated/control median peak-width ratios sit near 1, unlike the HA degron arms' 1.92x. The bin measure's failure is **triggered by near-total signal collapse**, which only a degron produces. Cycle 117's effects are in the 20-60% range, where the measure behaves.

Go deeper
The rule this establishes for the corpus

A bin-overlap retention figure is safe in the middle and unreliable at the extreme. Concretely: where the treated arm retains roughly 20-80% and its median peak width is within ~20% of the control's, the two measures agree. Where retention approaches zero — which is exactly where the most dramatic claims live — the bin measure inflates.

So the corpus does not need a mass re-scoring. It needs the treated/control peak-width ratio printed beside any retention figure, as a one-line trigger for when a re-score is required.

Facts and figures
scriptanalyses/what-leaves-when-the-fusion-is-degraded/run.py
statusrun
result summaryCycle 117's verdict holds under the corrected measure. MLL1: EPZ 88.3%, VTP 45.7%, both 17.9% retained; menin/DOT1L loss ratio 4.64 (was 3.33). H3K27ac control barely moves under either measure (EPZ 100.0%, VTP 99.6%, both 99.3%). The MOLM13 arms show no peak broadening, which is why the metric defect did not reach them.
preregistrationanalyses/what-leaves-when-the-fusion-is-degraded/PREREGISTRATION.md
inputs
[
  "GSE173599 GSM5271717, GSM5271718, GSM5271719, GSM5271720 (MOLM13 MLL1)",
  "GSE173599 GSM5271712-GSM5271715 (MOLM13 H3K27ac)"
]
How this could bite us back
  • A verdict that strengthens under a corrected method is reassuring but is not independent confirmation - it is the same four libraries scored differently. The cycle-117 claim still rests on n = 1 per arm in a non-index cell line, and neither this recheck nor the original can measure a noise floor.
Still open
  1. How many other retention figures in this corpus come from arms whose peak widths broadened? A single pass over the deposited peak tables computing the treated/control median-width ratio would triage the whole corpus cheaply.
  2. Does the menin/DOT1L separation hold at the 5 uM / 72 h DOT1L exposure the original MLL1-channel work used? Cycle 116 established no public deposit carries a menin or MLL1 readout at that dose, and this recheck does not change that.
Every source behind this entry (1)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    Re-scoring the MOLM13 MLL1 2x2 from GSE173599 under summit containment gives EPZ 88.3%, VTP 45.7% and combination 17.9% retention, a menin/DOT1L loss ratio of 4.64 against 3.33 under bin overlap.

    What this evidence cannot show

    n = 1 per arm, no replicates and no noise floor for these libraries. MOLM13 is KMT2A::MLLT3, not the index fusion KMT2A::AFF1, so both the original verdict and this recheck are partner-imported. The recheck changes only the scoring of the same four libraries - it is not an independent replication, and a defect in the underlying libraries would affect both measures identically. The ratio increase from 3.33 to 4.64 is arithmetic on small denominators (a loss of 11.7% in the DOT1L arm) and should not be read as a more precise estimate.

    Evidence that argues against this (1)
    • The strict measure could in principle inflate a ratio rather than correct it: taking proportionally more from the smaller effect is what produces 4.64 from 3.33, and a measure that systematically penalises weak effects would manufacture separation between a weak and a strong drug. The agreement to within 6 points on every arm argues against this, but the ratio - not the retention figures - is the quantity to distrust. Oncofusion knowledge graph 2026
Connected to