Oncofusion

start / graph / ana-q158-partner-exclusivity-does-discriminate-and-the-one-rival-that-fired-is-a-lineage-artefact

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

Partner-exclusivity does discriminate: of four comparator partner groups only AFDN fires, it is 4-of-4 AML, and lineage-matching collapses it to exactly chance while AFF1's class survives the same restriction

also called q158 pseudo-index test · the statistic can tell partners apart · AFDN's class is lineage not partner · 13 against 13.06

The worry was that this project's way of finding genes unique to one fusion partner might flag any group of patients at all, in which case none of its results would mean anything. Testing every other partner group shows it does not - only one rival group scored, and that one turned out to be explained by the patients all having myeloid rather than lymphoid leukaemia.

The same thing, for a scientist

Q154 named this as the cheapest experiment that could invalidate the whole q145-q157 line: *"run the identical permutation with the 6 KMT2A-ENL or 4 KMT2A-AFDN samples as the pseudo-index. If a class of ~six appears for any partner group of that size, partner-exclusivity has no purchase in this deposit at any scale."* **It had never been run at any scale. It has now been run for every partner group in the deposit.** 🟢 **Positive control:** a clean reimplementation rebuilds AFF1's published six-member class exactly (`ELF1, GNAQ, H3C15, IRX1, PAN3, SPAG7`) before any comparator is scored; the script exits non-zero otherwise [E1]. | partner group | n | k>=| observed | null | null 95th | p_upper | |---|---:|---:|---:|---:|---:|---:| | **KMT2A-AFF1** | 8 | 3 | **6** | 0.87 | 3 | **0.0029** | | KMT2A-ENL | 6 | 2 | 2 | 3.62 | 10 | 0.789 *(below chance)* | | **KMT2A-AFDN** | 4 | 2 | **6** | 1.39 | 5 | **0.0398** | | KMT2A-AF10 | 3 | 2 | 1 | 0.67 | 2 | 0.427 | | KMT2A-AF9 | 3 | 2 | 1 | 0.68 | 2 | 0.429 | | KMT2A-EPS15 | 2 | - | - | - | - | *not tested, n<3* | 🔵 **The feared outcome does not occur.** The largest comparator, **ENL at n=6, produces 2 against a null of 3.62 - below chance.** AF10 and AF9 sit at chance. **A coherent real partner group does not, by itself, manufacture an exclusive class** [E1]. 🔴 **One rival fired, and it is a lineage artefact.** AFDN gives **6 against 1.39 (p 0.0398)** - and **all four AFDN samples are AML**, while 16 of the 26 samples are B-lineage. Its members are myeloid and homeobox flavoured: `CLEC9A` (a myeloid/dendritic lectin), `IRX3`, `NKX2-3`, `PITX1`, `MPDU1`, `POLE3`. **Restricting the comparator pool to the 9 AML samples collapses it to 13 observed against a null of 13.06, p = 0.452** - chance to two decimal places [E2]. 🟢 **AFF1's class survives the same restriction.** Taking only its 7 B-lineage samples against the 16-sample B-lineage pool gives **7 against 1.66, p = 0.0044** [E2]. **The one test that dissolves the rival leaves the index arm intact**, which is the strongest form this control could have taken. 🟠 **But IRX1 is not in the lineage-matched class.** The B-only members are `ELF1, GNAQ, H3C15, HSPA2, MIDN, PAN3, SPAG7`. **IRX1 drops out**, because one of its three samples (`p318`) is the AML one, leaving 2 of 7 - below the k>=3 recurrence bar. **IRX1's three samples span two lineages, so the gene cannot appear in a lineage-matched class at this threshold.** The class is partner-specific; IRX1's membership in it is not lineage-robust [E2].

Go deeper
What this licenses, and the multiple-testing arithmetic

🟢 Licensed: the exclusive-recurrent statistic discriminates between partner groups in this deposit. Q154's invalidating scenario - every coherent group producing a class - is refuted by ENL, AF10 and AF9. The q145-q157 line is not resting on a statistic with no discrimination.

🟠 Four groups were tested, so the threshold moves. A Bonferroni correction across the four gives 0.0125. AFF1 at 0.0029 survives it; AFDN at 0.0398 does not - so even before the lineage test, AFDN was the weaker call. The lineage test then removes it entirely.

🔴 The comparison is not apples-to-apples and I chose the scaling. AFF1 is tested at n=8, k>=3; the smaller groups are tested at k>=2 because k>=3 of 3 or 4 samples is a different demand. I set kmin = max(2, round(k * 3/8)). A different scaling would change the comparator numbers, though not the AFDN lineage result, which is computed at fixed k>=2 on both sides of the restriction.

🔵 The AFDN collapse is the cleanest number in the analysis and it should be read as the headline rather than AFF1's survival: 13 observed against a null of 13.06. A confound that disappears to two decimal places under the matched comparison is a confound identified, not merely suspected.

Facts and figures
scriptanalyses/q158-within-family-pseudo-index/run.py
addendum scriptanalyses/q158-within-family-pseudo-index/lineage_check.py
statusrun
n perm100000
seed20260924
answers question fromQ154 hand-back, cheapest remaining gap
bonferroni threshold 4 tests0.0125
verdictthe statistic discriminates; the one rival that fired is a lineage artefact; IRX1 itself is not lineage-robust
groups
KMT2A-EPS152.000KMT2A-EPS15: 2KMT2A-AF103.000KMT2A-AF10: 3KMT2A-AF93.000KMT2A-AF9: 3KMT2A-AFDN4.000KMT2A-AFDN: 4KMT2A-ENL6.000KMT2A-ENL: 6KMT2A-AFF18.000KMT2A-AFF1: 8
positive control
{
  "class": [
    "ELF1",
    "GNAQ",
    "H3C15",
    "IRX1",
    "PAN3",
    "SPAG7"
  ],
  "reproduced": true,
  "aborts_if_failed": true
}
groups — the numbers behind the chart
{
  "KMT2A-AFF1": {
    "n": 8,
    "kmin": 3,
    "obs": 6,
    "null": 0.87,
    "p_upper": 0.00292
  },
  "KMT2A-ENL": {
    "n": 6,
    "kmin": 2,
    "obs": 2,
    "null": 3.62,
    "p_upper": 0.78807
  },
  "KMT2A-AFDN": {
    "n": 4,
    "kmin": 2,
    "obs": 6,
    "null": 1.39,
    "p_upper": 0.03984,
    "members": [
      "CLEC9A",
      "IRX3",
      "MPDU1",
      "NKX2-3",
      "PITX1",
      "POLE3"
    ],
    "all_AML": true
  },
  "KMT2A-AF10": {
    "n": 3,
    "kmin": 2,
    "obs": 1,
    "null": 0.67,
    "p_upper": 0.42724
  },
  "KMT2A-AF9": {
    "n": 3,
    "kmin": 2,
    "obs": 1,
    "null": 0.68,
    "p_upper": 0.42911
  },
  "KMT2A-EPS15": {
    "n": 2,
    "tested": false
  }
}
lineage matched
{
  "AFDN_vs_AML_only": {
    "pool": 9,
    "k": 4,
    "obs": 13,
    "null": 13.06,
    "p_upper": 0.45221
  },
  "AFF1_B_vs_B_only": {
    "pool": 16,
    "k": 7,
    "obs": 7,
    "null": 1.66,
    "p_upper": 0.00442,
    "members": [
      "ELF1",
      "GNAQ",
      "H3C15",
      "HSPA2",
      "MIDN",
      "PAN3",
      "SPAG7"
    ],
    "IRX1_present": false,
    "why_IRX1_absent": "p318 is AML, so IRX1 has 2 of 7 B-lineage AFF1 samples, below k>=3"
  }
}
measurements
[
  {
    "quantity": "comparator partner groups producing an exclusive-recurrent class above chance",
    "value": 1,
    "unit": "of 4 tested",
    "subject": "KMT2A-AFDN at p_upper 0.0398; ENL below chance, AF10 and AF9 at chance",
    "universe": "4 groups with n>=3; Bonferroni threshold 0.0125, which AFDN does not clear and AFF1 (0.0029) does"
  },
  {
    "quantity": "AFDN's exclusive class against a lineage-matched AML-only comparator pool",
    "value": 13.06,
    "unit": "null mean against 13 observed",
    "subject": "p_upper 0.452 - the class disappears to two decimal places once lineage is held fixed",
    "universe": "AFF1's class under the same restriction gives 7 against 1.66, p_upper 0.0044, and survives"
  }
]
Still open
  1. Re-run the comparator groups under a pre-registered constant recurrence fraction rather than my kmin = max(2, round(k*3/8)) scaling. The AFDN lineage result is robust to this because it is computed at fixed k>=2 on both sides, but the cross-group table is not.
  2. Decide what IRX1's absence from the lineage-matched class means for the entries that treat it as the partner-specific address. On the most controlled version of this test IRX1 cannot appear, and that should be stated in gene-irx1 and in the q147 headline rather than left here.
  3. Test whether AFDN's six genes are simply AML-expressed. CLEC9A, IRX3, NKX2-3 and PITX1 are myeloid or developmental; checking them against an AML expression reference would convert 'lineage artefact' from an inference about the permutation into a statement about the genes.
  4. Separate lineage from partner properly, which this deposit may not permit. KMT2A-AFDN is predominantly myeloid and KMT2A::AFF1 predominantly lymphoid, so the confound is structural; a deposit with lineage-discordant cases of both partners would be needed.
Every source behind this entry (2)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    Running the identical gene-body exclusive-recurrent statistic and label-permutation null for every partner group in GSE252378 with at least three non-drug-exposed samples gives: KMT2A-AFF1 (n=8, k>=3) 6 observed against a null of 0.87, p_upper 0.0029; KMT2A-ENL (n=6, k>=2) 2 against 3.62, p_upper 0.789, below chance; KMT2A-AFDN (n=4, k>=2) 6 against 1.39, p_upper 0.0398; KMT2A-AF10 (n=3, k>=2) 1 against 0.67, p_upper 0.427; KMT2A-AF9 (n=3, k>=2) 1 against 0.68, p_upper 0.429. KMT2A-EPS15 has only 2 samples and was not tested. The run is gated on a positive control that rebuilds AFF1's published six-member class exactly.

    PMID 39472576DOI
    What this evidence cannot show

    The recurrence threshold is not constant across groups: AFF1 is scored at k>=3 of 8 while the smaller groups are scored at k>=2, using kmin = max(2, round(k*3/8)), which is my scaling choice and not pre-registered. A different scaling would move the comparator numbers. Four groups were tested, so a Bonferroni threshold is 0.0125 - AFF1 clears it, AFDN does not. EPS15 (n=2) is untested, so 'every partner group' means every group with n>=3. The partner annotation is taken from this project's own q147 axis rather than re-read from GEO this session.

    Evidence that argues against this (2)
    • The comparator groups are smaller than the index group, and a smaller group has fewer chances to accumulate a recurrent class - so some of AFF1's advantage may be statistical power rather than biology. The permutation null is computed at each group's own k, which controls for this in expectation, but it cannot restore the resolution a 3-sample group does not have. ENL at n=6 is the only comparator with real power and it came in below chance, which is the strongest single line here.
    • A group scoring at chance is not evidence that its partner has no exclusive targets - only that this instrument cannot find them at this depth with this many samples. Reading ENL's 2-against-3.62 as 'ENL has no exclusive class' would be the same error as treating a bounded absence as an established negative, which this corpus records repeatedly.
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.journal-article

    All four KMT2A-AFDN samples are AML while 16 of the 26 scored samples are B-lineage, and AFDN's nominally significant exclusive class (CLEC9A, IRX3, MPDU1, NKX2-3, PITX1, POLE3) collapses when the comparator pool is restricted to the 9 AML samples: 13 observed against a null of 13.06, p_upper 0.452. Applying the same restriction to the index arm, AFF1's 7 B-lineage samples against the 16-sample B-lineage pool give 7 observed against 1.66, p_upper 0.0044; the lineage-matched members are ELF1, GNAQ, H3C15, HSPA2, MIDN, PAN3 and SPAG7, and IRX1 is absent from them because one of its three samples, p318, is AML, leaving 2 of 7.

    PMID 39472576DOI
    What this evidence cannot show

    The AML pool is 9 samples of which 4 are the AFDN index, leaving only 5 comparators, so the lineage-matched AFDN test is badly underpowered and its p = 0.452 should be read as 'no evidence of an effect' rather than 'evidence of no effect'. The observed count rises from 6 to 13 under restriction simply because exclusivity is easier to achieve against 5 comparators than against 22 - the null rises with it, which is the point, but the raw counts are not comparable across the two rows. Lineage is taken from this project's stored LINEAGE map and was not re-derived from the deposit's metadata this session. The single T-lineage sample is excluded from both pools.

    Evidence that argues against this (2)
    • Lineage and partner are confounded in this deposit by design, not by accident: KMT2A-AFDN is a predominantly myeloid fusion and KMT2A::AFF1 a predominantly lymphoid one, so 'the AFDN class is lineage, not partner' and 'the AFDN class is partner, and partner determines lineage' fit these data equally well. This analysis cannot separate them, and the same objection applies in principle to AFF1's surviving class.
    • IRX1 dropping out of the lineage-matched class is at least as consistent with the class being an artefact as with IRX1 being unusual. If the gene that motivated the entire line cannot appear in the most controlled version of the test, the controlled test is not evidence for IRX1 - and the entries that treat IRX1 as the partner-specific address should say so.
Connected to