Oncofusion

start / graph / ana-q122-curated-partner-dependence-gets-its-species-independent-leg-and-the-culture-noise-is-a-third-of-the-effect

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.

CURATED from committed-but-unread results: murine MLL-AF9 (MLLT3) retains H3K79me2 like human MOLM13 (MLLT3) and unlike RS4;11 (AFF1) - four confirmations fire, two refutations do not, the cross-partner delta is 3.2-3.6x the between-culture spread, and the flatness survives an amplitude stratification in 8 of 8 measurements

also called the species-independent leg of partner dependence · is the H3K79me2 difference bigger than the culture noise · Q122 curation, criteria C1-C3, C6, C8, R1, R3, R5

Two leukaemia cell lines carrying different fusion partners had given very different answers about whether a chemical mark on DNA packaging survives a menin inhibitor, and the obvious worry was that the difference was just noise between cell cultures rather than anything to do with the partner. This checks that, and the difference between partners turns out to be three to four times larger than the variation between two cultures of the same cell type - and a mouse line carrying the same partner as one of the human lines behaves like that human line rather than the other, in a completely separate experiment. The result also survives being recomputed on only the strongest sites, in all eight measurements where that was tested, so it is not an artefact of averaging weak signal with strong.

The same thing, for a scientist

**PROVENANCE: coordinator curation of committed-but-unread results.** Q122 was killed by the account session limit before writing entries. Cycle 82 salvaged two results from its `results.json`; **six further scored criteria sat measured and committed with no entry.** This curates them. Every figure is read from the beat's own file; nothing was re-derived from raw data by me. **The question.** Q120 found RS4;11 (KMT2A::AFF1) at **0.6578** and MOLM13 (KMT2A::MLLT3) at **0.8667** for H3K79me2 retention under VTP-50469 — a 0.2089 gap, while nine other cross-line comparisons agreed within 0.04. `hyp-q120` proposed the fusion partner sets it. The obvious alternative is that 0.2089 is culture-to-culture noise. **Four confirmations fire; two refutations do not.** | criterion | population | observed | threshold | verdict | |---|---|---|---|---| | **C1** species-independent leg | murine MLL-AF9, S_FLAG, both cultures | V = **0.7734**, **0.8318** | ≥ 0.7623 | **FIRES** | | **C2** at the mark's own domains | murine MLL-AF9, S_K79 | V = **0.9117**, **0.8462** | ≥ 0.7623 | **FIRES** | | **C3** spread < cross-partner delta | F3 vs F3_NT | **0.0584**, **0.0655** | < 0.2089 | **FIRES** | | **C6** independent-series replication | human MOLM13, GSE260742 | **0.9063** | ≥ 0.7623 | **FIRES** | | **R1** opposite answer across species | needs V ≤ 0.7078 in **both** cultures | 0.7734, 0.8318 | — | **DOES NOT FIRE** | | **R3** delta inside culture spread | fires if C3 fails | C3 held | — | **DOES NOT FIRE** | | **R5** instrument | any of I0/I1/I2 failing | no failures | — | **DOES NOT FIRE** | **The number that does the most work is C3.** Between two cultures of the same murine line the spread is **0.0584 and 0.0655**; the cross-partner delta is **0.2089**. **The partner effect is 3.2 to 3.6 times the culture-to-culture variation.** That is the comparison that separates a finding from noise, and it had never been made for this mark. **C6 is an independent-series replication, which this corpus rarely gets.** MOLM13 reads **0.9063 in GSE260742** against q120's **0.8667 in GSE127507** — different deposit, different submitters' processing, different timepoint (96 h vs 72 h), same partner, same answer. **And C8 kills the aggregation-artefact reading in 8 of 8 measurements.** Recomputing each share ratio on the **top decile of sites** instead of the whole set moves it by **0.0029 to 0.0371**, all inside the 0.10 band declared in advance — including the two most load-bearing figures (K562 VTP 0.8291 → 0.8361; MOLM13 VTP 0.9063 → 0.8721). Q120 had flagged exactly this as the cheapest thing that could dismantle its own conclusion. It does not.

Go deeper
The comparison mixes two series, and the threshold is derived from the values it discriminates

Two design features need stating because both are the kind of thing that reads as rigour until examined.

The cross-partner delta of 0.2089 comes from GSE127507 (Q118's RS4;11 and Q120's MOLM13) while the murine arms come from GSE260742. So C1 and C2 compare a murine measurement in one deposit against a threshold built from human measurements in another, with different submitters, different processing and different timepoints. C6 is what partially rescues this — it re-measures MOLM13 inside GSE260742 and gets 0.9063, close to the 0.8667 the threshold was built from, which shows the two series are not wildly offset for this quantity. Without C6, C1 and C2 would be uninterpretable.

The 0.7623 threshold is the midpoint of the two values it discriminates between (0.6578 and 0.8667). That is a legitimate design for a two-way discriminator — it asks "which of two known answers is this closer to" — but it is not a test of whether the murine value is high in absolute terms, and it should never be quoted as one. A murine value of 0.77 clears it while still being well below MOLM13's 0.87.

F3 and F3_NT are two cultures, not replicates. Q120 established that nothing in GSE260742's 158 sample records says they are one culture or two arms of one experiment. So C3's 0.0584/0.0655 is a between-culture spread, which is the right comparator for this purpose and is probably conservative — genuine technical replicates would spread less.

Facts and figures
CURATION PROVENANCEQ122 computed these and was killed by the account session limit before writing any node file. Cycle 82 salvaged two results from its results.json; these six criteria were measured, committed and uncurated until cycle 83. No figure re-derived from raw data by the coordinator.
the number that does the most workC3. Between-culture spread 0.0584/0.0655 against a cross-partner delta of 0.2089 - the partner effect is 3.2-3.6x the culture variation, and that comparison had never been made for this mark.
the confound this cannot touchLINEAGE tracks the partner perfectly across every system compared: RS4;11 is B-lymphoblastic, MOLM13 and murine MLL-AF9 are myeloid. A partner-matched pair in one lineage, or a lineage-matched pair with different partners, is the discriminating design and does not exist in these deposits.
absence classmeasured-not-curated
absence class noteThese six criteria were computed, scored and committed to this repository by Q122 and sat with no entry for one cycle because the beat died before curating them. This entry curates them. Strictly measured-not-curated: the measurement was not merely public, it was already in the repo's own analyses directory.
C1 species independent leg
{
  "V_F3": 0.7734,
  "V_F3NT": 0.8318,
  "threshold": 0.7623,
  "verdict": "FIRES"
}
C2 at the marks own domains
{
  "V_F3": 0.9117,
  "V_F3NT": 0.8462,
  "threshold": 0.7623,
  "verdict": "FIRES"
}
C3 spread vs cross partner delta
{
  "abs_diff_at_S_FLAG": 0.0584,
  "abs_diff_at_S_K79": 0.0655,
  "cross_partner_delta": 0.2089,
  "ratio": "3.2x to 3.6x",
  "verdict": "FIRES"
}
C6 independent series replication
{
  "V_MOLM13_GSE260742": 0.9063,
  "V_MOLM13_GSE127507_q120": 0.8667,
  "verdict": "FIRES"
}
C8 flatness is not an aggregation artefact
{
  "band": 0.1,
  "n_measurements": 8,
  "n_firing": 8,
  "abs_gap_range": "0.0029 to 0.0371",
  "most_load_bearing": {
    "K562_VTP": "0.8291 -> 0.8361",
    "MOLM13_VTP": "0.9063 -> 0.8721"
  },
  "verdict": "FIRES in 8 of 8"
}
R1 opposite answer across species
{
  "fires_if": "V(S_FLAG) <= 0.7078 in BOTH cultures",
  "observed": [
    0.7734,
    0.8318
  ],
  "verdict": "DOES NOT FIRE"
}
R3 delta inside culture spread
{
  "fires_if": "C3 fails",
  "verdict": "DOES NOT FIRE"
}
R5 instrument
{
  "I0_failures": [],
  "I2_failures": [],
  "verdict": "DOES NOT FIRE"
}
Still open
  1. Rename or re-scope hyp-q120. Lineage tracks the fusion partner perfectly in every system compared here, so 'partner dependence' and 'lineage dependence' are indistinguishable on this evidence and the hypothesis title claims the former.
  2. A finer amplitude stratification than top-decile-versus-whole-set. Q120's own stratification moved a different channel from 0.8305 to 0.5304 at the strongest tenth, so a decile cut can move figures substantially in this data; quartiles across the full range would be a stronger test of C8.
  3. Is there any deposit anywhere with a partner-matched pair in one lineage - e.g. KMT2A::AFF1 and KMT2A::MLLT3 both in B-lineage, or both in myeloid? That single comparison would settle the partner-versus-lineage question that all of this rests on.
Every source behind this entry (1)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.analysis

    In GSE260742 murine MLL-AF9 (KMT2A::MLLT3), H3K79me2 within-sample share ratios under 500 nM VTP-50469 are V = 0.7734 and 0.8318 at the FLAG-defined site set and 0.9117 and 0.8462 at the mark's own domains across two culture families, all above the 0.7623 midpoint discriminating Q118's RS4;11 (0.6578) from Q120's MOLM13 (0.8667). The between-culture spread is 0.0584 and 0.0655 against a cross-partner delta of 0.2089, a ratio of 3.2 to 3.6. Human MOLM13 re-measured inside GSE260742 gives 0.9063 against 0.8667 in GSE127507. Recomputing on the top decile of sites rather than the whole set moves every one of eight measurements by 0.0029 to 0.0371, inside a 0.10 band declared in advance. Pre-registered refutations R1, R3 and R5 do not fire.

    analyses/q122-h3k79me2-three-arms-with-replicates/results.json, criteria C1, C2, C3, C6, C8, R1, R3 and R5; pre-registration written before any bin file existed; this repository · Oncofusion knowledge graph, Q122 beat (computation) and coordinator (curation of the numbers into this entry) · internal analysis, reproducible from the repository 2026
    What this evidence cannot show

    CONFIDENCE IS moderate FOR PROVENANCE REASONS FIRST: the beat that computed these figures died before writing any entry, no agent ever wrote prose against them, and I re-derived nothing from raw data - every value is read out of its results.json. DESIGN LIMITS, all real: (1) C1 and C2 compare a GSE260742 murine measurement against a threshold built from GSE127507 human measurements - different deposits, submitters, processing and timepoints - and it is C6 that partially rescues this by re-measuring MOLM13 inside GSE260742; (2) the 0.7623 threshold is the MIDPOINT of the two values it discriminates, so clearing it means 'closer to MOLM13 than to RS4;11' and NOT 'high in absolute terms'; (3) F3 and F3_NT are two cultures, not replicates, and nothing in 158 sample records establishes their relationship; (4) no usable spike-in in this series, so every figure is a within-sample share ratio. Doses and durations differ across the compared arms (48 h murine, 96 h MOLM13 GSE260742, 72 h MOLM13 GSE127507).

    Evidence that argues against this (3)
    • The strongest objection is that 'partner dependence' is not the only variable that separates RS4;11 from MOLM13 and the murine line. RS4;11 is a B-lymphoblastic line and MOLM13 and murine MLL-AF9 are myeloid - so LINEAGE tracks the partner perfectly across every system compared here, and nothing in this analysis can separate the two. A partner-matched pair in the same lineage, or a lineage-matched pair with different partners, is the discriminating design and does not exist in these deposits. `hyp-q120` should arguably be renamed to name the confound.
    • C3's reassurance is narrower than it looks. It compares between-CULTURE spread within one murine line against a cross-partner delta measured between two human lines in a different series. A within-line spread in mouse does not bound the between-line, between-series, between-submitter variation that the 0.2089 delta actually spans. The honest version is that culture noise is too small to explain the delta IF all other sources of variation are comparable to culture noise, which is assumed and not shown.
    • C8's stratification tests only the top decile against the whole set. Q120's own amplitude stratification found MLL1-N falling from 0.8305 to 0.5304 at the strongest tenth for a different channel, so a decile cut CAN move a figure substantially in this data - that it does not here for H3K79me2 is informative, but a finer stratification (quartiles or deciles across the whole range) would be a stronger test and was not run.
Connected to