Oncofusion

start / graph / ana-interactome5-lineage-split

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q5-interactome5

Lineage-split dependency for the round 5 panel: 23 genes, 2 myeloid-identity artefacts caught, and 3 survivors of which one has no literature at all

also called interactome5 lineage split · round 5 DepMap lineage split

The cell lines used to find KMT2A-specific vulnerabilities are mostly myeloid, so any gene that myeloid cells happen to need looks specific for free. This analysis scores every candidate twice - once in the myeloid lines and once in the B-cell lines - and reports both numbers.

The same thing, for a scientist

Every dependency number stated anywhere in interactome round 5 comes from this run. Twenty-three genes covering FACT, NSL, MSL, the COMPASS remainder, Mediator and the MEN1/KMT2A/DOT1L reference axis were scored with the KMT2A-rearranged index arm **split by lineage**, each half expressed in standard deviations of the other-leukaemia arm, with Ensembl ids taken from a cached HGNC set and the API's `approvedSymbol` asserted against the requested symbol before any row was kept [E1]. **Two genes are flagged as myeloid-identity artefacts: CXXC1 and MED1.** Both look excellent on the single-number genome-wide scan - CXXC1 rank 414, MED1 rank 81 - and both collapse in the B-lineage half (CXXC1 myeloid z -0.97 against B-lineage -0.28; MED1 -0.86 against -0.13) [E1][E2]. **Three genes survive: MEN1, KMT2A and MED23.** MEN1 is near-symmetric (-2.29 and -2.28), KMT2A likewise (-1.48 and -1.22), and **MED23 is stronger in the B-lineage half than the myeloid half (-1.85 against -0.77)** - the opposite shape from the artefact class, on a pan-essentiality of only -0.2096, and with **zero published records** against MLL or KMT2A [E1][E3]. Everything else in the panel is 'weak or ambiguous'. NPM1 could not be scored: too few lines returned (3 myeloid, 1 lymphoid, 7 comparator) [E1].

Go deeper
The table, both halves, with raw effects where they matter

z is in standard deviations of the other-leukaemia arm. A z is not an effect size: it is scaled by the comparator SD, which is small for genes every leukaemia line needs, so pan-essentiality is given alongside 1.

genepanmyeloid zB-lineage zverdict
MEN1-0.362-2.29-2.28lineage-independent
KMT2A-0.206-1.48-1.22lineage-independent
MED23-0.206-0.77-1.85lineage-independent
MED24-0.443-0.96-0.42weak or ambiguous
MSL30.019-1.03-0.40weak or ambiguous
CXXC1-0.384-0.97-0.28MYELOID-IDENTITY ARTEFACT
MED1-0.753-0.86-0.13MYELOID-IDENTITY ARTEFACT
SETD1B-0.241-0.65+0.18weak or ambiguous
MED12-0.557-0.60-0.04weak or ambiguous
MED14-1.127-0.59-0.81weak or ambiguous
DOT1L-0.147-0.53-0.88weak or ambiguous
MSL1-0.365-0.27-0.42weak or ambiguous
MCRS1-0.622-0.30+0.42weak or ambiguous
SETD1A-0.910-0.12-0.73weak or ambiguous
KAT8-1.453-0.13+0.24weak or ambiguous
SUPT16H-1.760-0.11+0.86weak or ambiguous
KANSL3-1.183-0.01+0.25weak or ambiguous
OGT-1.466-0.01-0.43weak or ambiguous
MSL2-0.083+0.09-0.09weak or ambiguous
KANSL1-0.738+0.16+0.36weak or ambiguous
PHF20-0.079+0.35-0.33weak or ambiguous
KANSL2-0.982+0.41+1.32weak or ambiguous
SSRP1-1.272+0.54+0.90weak or ambiguous

Read the MEN1 row as the calibration. A genuine, lineage-independent dependency looks like -2.29 and -2.28: two large numbers that agree. Nothing else in this panel looks like that. MED23's halves differ by 1.08, so its 'lineage-independent' label is a threshold decision rather than an observed symmetry, and this analysis says so rather than letting the label stand unqualified.

What the two artefacts cost, and what catching them saved

CXXC1 was deferred by round 4 with the note that 'rank 414 (-0.161) is a real signal'. It is a real number and it is a myeloid-weighted one. Had this beat written CXXC1 on the strength of the genome-wide rank, the graph would carry a B-ALL lead whose B-lineage z is -0.28.

MED1 was in the brief's priority list, on the reasonable ground that MED1 is a well-known Mediator subunit and ranks 81 of 17,728. Its B-lineage z is -0.13. Its PubMed literature against KMT2A is also empty on title-check. Both halves of the case dissolved at once.

The saving is not only in avoiding two bad entries. It is that both genes are now recorded with both numbers, so a future cycle looking at rank 414 or rank 81 finds the split attached rather than rediscovering the rank and re-making the error. That is the whole point of writing the artefacts down instead of silently dropping them.

The cost of the method is honest too: four B-lineage index lines. A single line can move a mean, so 'artefact' here means 'does not replicate in the B-lineage half of a small panel', not 'demonstrated to be lineage-driven'. The alternative reading - that CXXC1 is genuinely required in KMT2A-rearranged AML and not in KMT2A-rearranged B-ALL - is biologically reasonable and is not excluded.

The MED23 problem

MED23 is the only gene in this panel that is simultaneously (a) strongly selective on the genome-wide scan, rank 16 of 17,728, (b) not floor-compressed, with a pan-essentiality of only -0.2096, (c) stronger in the B-lineage half, which is the index disease's half, and (d) entirely unpublished in connection with MLL or KMT2A 3.

Each of (a) to (c) is a reason to take it seriously. (d) is a reason not to. A genome-wide top-20 hit in a disease this heavily studied, with no literature whatsoever, is more often an artefact than a discovery - and the Mediator panel it belongs to contains a demonstrated artefact (MED1) and ranks improbably well as a whole (MED23 16, MED1 81, MED24 128, MED12 131, MED14 453 2).

Three cheap analyses would separate the possibilities and none was run in this cycle: printing the four B-lineage per-line values, running the permutation and lineage-conditioning scripts that already exist in this repository, and checking whether MED23 expression tracks KMT2A rearrangement. That omission is recorded here as the largest methodological shortcut of interactome round 5.

Facts and figures
scriptanalyses/depmap-interactome5-lineage-split/run.py
statusrun
result summary23 genes scored with the KMT2A-rearranged index arm split by lineage. 2 myeloid-identity artefacts (CXXC1, MED1); 3 lineage-independent (MEN1, KMT2A, MED23); 18 weak or ambiguous; NPM1 unscorable for want of lines.
n genes23
calibration noteMEN1 at myeloid z -2.29 and B-lineage z -2.28 is what a genuine lineage-independent dependency looks like: two large numbers that agree. MED23's halves differ by 1.08, so its label is a threshold decision rather than observed symmetry.
caveatEight myeloid and five B-lineage index lines available; six and four contributed for most genes, so a single line moves a mean. Split is by lineage LABEL, not expression signature. Pan-essential genes sit near the Chronos floor where the z is a difference between two lethal effects. Bounds lineage specificity only; says nothing about therapeutic window.
inputs
[
  "DepMap Chronos gene effect via Open Targets GraphQL",
  "cached HGNC complete set for Ensembl id resolution"
]
artefacts
[
  "CXXC1",
  "MED1"
]
survivors
[
  "MEN1",
  "KMT2A",
  "MED23"
]
unscorable
[
  "NPM1"
]
analyses not run
[
  "permutation test",
  "lineage-conditioned analysis",
  "line-normalised analysis",
  "per-line value inspection for MED23",
  "expression normalisation"
]
How this could bite us back
  • This analysis bounds lineage specificity and nothing else. A gene that survives the split is not thereby a target, and MEN1 - which survives most convincingly - is already drugged.
  • The verdict labels are this run's own thresholds. Treating 'MYELOID-IDENTITY ARTEFACT' as a demonstrated mechanism rather than a failure to replicate in four B-lineage lines would overstate what was done.
  • Applying the split only to a 23-gene panel means the rest of the graph's dependency claims remain unsplit. The rule that every DepMap claim must be lineage-split is only as good as its coverage, and coverage is currently partial.
Still open
  1. Do the four B-lineage lines agree on MED23, or does one carry the z of -1.85? Minutes of work, not done.
  2. Why does the whole Mediator panel rank inside the top 460? An expression-normalised rerun would say whether this is transcriptional output rather than fusion biology.
  3. Is 'myeloid-identity artefact' distinguishable from 'genuinely AML-restricted dependency' with four B-lineage lines? Probably not, and the label should be read as the weaker of the two claims.
  4. Could NPM1 be scored at all with a larger comparator set, given that the NPM1c line OCI-AML3 is the most menin-dependent line in this project's data?
Every source behind this entry (3)
  1. E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    For the round 5 panel of 23 genes, DepMap Chronos gene effects were retrieved through the Open Targets GraphQL API for leukaemia lines only, the KMT2A-rearranged index arm was split into myeloid and B-lineage halves, and each half was scored in standard deviations of the other-leukaemia arm. Ensembl ids were taken from a cached HGNC complete set and the API's approvedSymbol was asserted against the requested symbol before any row was kept. Two genes were classified myeloid-identity artefacts (CXXC1, MED1) and three lineage-independent (MED23, MEN1, KMT2A). NPM1 could not be scored: too few lines (3 myeloid, 1 lymphoid, 7 comparator). Reference values: MEN1 myeloid z -2.29 and B-lineage z -2.28; KMT2A -1.48 and -1.22; MED23 -0.77 and -1.85.

    Interactome round 5 panel, lineage-split DepMap dependency (FACT, NSL, MSL, COMPASS remainder, Mediator) · Broad Institute DepMap / Open Targets Platform; analyses/depmap-interactome5-lineage-split 2026
    What this evidence cannot show

    All rows read out of analyses/depmap-interactome5-lineage-split/results.json on 2026-09-19 in a scripted pass; no number in this entry was typed from recall. **Eight myeloid and five B-lineage index lines are available in principle and six and four contributed for most genes, so a single line moves a mean.** The split is by lineage LABEL, not by expression signature, so a B-lineage line with a myeloid expression programme is mis-assigned. Pan-essential genes sit near the Chronos floor where a z is a difference between two lethal effects and compression is severe. **This bounds lineage specificity only; it is not evidence of a therapeutic window.** The verdict labels come from this run's own thresholds and are not independent demonstrations.

    Evidence that argues against this (1)
  2. E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    In the project's genome-wide KMT2A-rearranged selectivity scan of 17,728 genes, the two genes this run flags as lineage artefacts rank highly on the single-number statistic: MED1 rank 81 (difference -0.3114, pan -0.7573) and CXXC1 rank 414 (-0.1606, pan -0.3833). The Mediator panel as a whole ranks improbably well: MED23 16 (-0.4827), MED1 81, MED24 128 (-0.2698), MED12 131 (-0.268), MED14 453 (-0.1541).

    Genome-wide KMT2A-rearranged selectivity scan, 17,728 genes (DepMap Chronos via Open Targets) · Broad Institute DepMap / Open Targets Platform; analyses/depmap-kmt2a-selective-scan-full 2026
    What this evidence cannot show

    Rows read out of analyses/depmap-kmt2a-selective-scan-full/results.json on 2026-09-19 in a scripted pass. Eleven KMT2A-rearranged against 43 other leukaemia lines, AML-heavy, none infant-derived. The observation that five Mediator subunits fall inside the top 460 is a pattern, not a test; with a complex of that size some clustering is expected by chance, and no formal enrichment statistic was computed.

    Evidence that argues against this (1)
  3. E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    As of 2026-09-19, '(MLL OR KMT2A) AND (MED23 OR SUR2 OR CRSP133 OR DRIP130 OR ARC130)' returns 0 PubMed records, and '(MED23 OR SUR2) AND (leukemia OR leukaemia OR AML)' returns 4, none of which is a MED23 leukaemia study on title-check.

    What this evidence cannot show

    Counts from the esearch `count` field, queried 2026-09-19, with all four records in the second query title-checked. **The symbol set omitted CRSP3 and KIAA1216**, two UniProt-listed synonyms, so this zero is a floor with a known hole in it. The alias SUR2 collides with a C. elegans gene and inflates rather than improves recall.

    Evidence that argues against this (1)
Connected to