Oncofusion

start / graph / concept-expressed-required-maintains-three-evidence-classes

Explainerstrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.Q60-cellstate2

Expressed, required, maintains — three different claims about where a gene matters, and the graph must label which one it has

also called expression is not dependency is not function · evidence class for a cell-type claim · cell type versus cell state axes

Saying a gene is switched on in a cell is not the same as saying the cell needs it, and neither is the same as saying the gene holds the cell in a particular condition. These are three different strengths of claim and most of what is known is only the weakest one.

The same thing, for a scientist

A cell-context claim about a gene falls into one of three evidence classes that are routinely conflated: (1) EXPRESSED — transcript or protein detected in a cell type (HPA, GTEx, single-cell atlases); (2) REQUIRED — loss of the gene impairs that cell type (DepMap lineage-split gene effect, conditional knockout, competitive transplantation); (3) MAINTAINS A STATE — the gene sustains or enforces a defined cellular condition such as quiescence, a differentiation block, drug tolerance or a retained fetal programme. Class 1 is weak and is usually the only class available. Class 3 is the most interesting and the least evidenced. Separately, CELL TYPE (what kind of cell: pro-B, GMP, HSC) and CELL STATE (what condition it is in: quiescent, blocked, persisting) are orthogonal axes that this graph keeps as separate node families, and a gene can be type-restricted and state-agnostic or the reverse.

Go deeper
Why this is not pedantry in this project

This repository has already been bitten twice by class confusion in a way that changed a conclusion.

The lineage confound. A genome-wide scan contrasting KMT2A-rearranged lines against other leukaemia lines ranked CEBPA 25th and SPI1 43rd. Both are artefacts of the index panel being myeloid-enriched: split by lineage, CEBPA moves +0.15 and SPI1 +0.92 in the B-lineage half. A class-2 statement ('required in KMT2A-rearranged leukaemia') was really a class-1 statement about myeloid identity. 1

The floor. Six of the sixteen highest-degree core genes in this graph — WDR5, KAT8, RBBP5, BRD4, ELL, PAF1 — are depleted in over 93% of all 1,258 screened cell lines. For those, class 2 is satisfied everywhere and therefore carries no cell-type information at all. 2

The classes, with what each can and cannot support
classtypical sourcesupportsdoes NOT support
1. EXPRESSEDHPA RNA/protein, GTEx bulk TPM, single-cell atlasthat the protein is presentanything about requirement; GTEx is bulk tissue with no cell-type resolution within a tissue
2. REQUIREDDepMap gene effect split by lineage, conditional KO, competitive transplantthat removing it impairs that cell typewhich state within the cell type; and nothing at all if the gene is floor-compressed
3. MAINTAINS A STATEinducible KO with state readout, transplant serial limiting dilution, sorted-population profilingthat the gene holds the cell in a conditionthat the state is the therapeutic target rather than a correlate

And a fourth thing that is not a class: a gene being a member of a complex whose complex was studied. Brief rule 64 exists because subunit-level absences are routinely false for that reason.

The cleanest worked example in this graph

PSIP1/LEDGF is class 2 and class 2 only, and it is unusually clean: conditional Psip1 knockout mice have essentially normal haematopoiesis, while MLL-rearranged leukemogenesis fails without it. 3 The claim 'LEDGF matters in blood' is false; the claim 'LEDGF is required in the leukaemic cell and not in the normal one' is true and is the entire therapeutic argument. The wild-type-versus-fusion-context distinction is therefore not a caveat here — it is the finding.

The mirror case is MLLT3/AF9, which governs human HSC self-renewal in normal haematopoiesis 4 and is required in essentially no cancer cell line (0.4% of 1,258 depleted below -0.5). 2 Reading the normal-cell function as a leukaemia dependency would invert the therapeutic argument.

Facts and figures
fourth considerationwhether the result was shown for the FUSION context or for the wild-type protein in normal cells - recorded per gene as data.cell_type.context
classes
expressed_in1.000expressed_in: 1required_in2.000required_in: 2maintains_state3.000maintains_state: 3
classes — the numbers behind the chart
[
  {
    "class": 1,
    "name": "expressed_in",
    "strength": "weak",
    "typical_sources": [
      "Human Protein Atlas RNA and protein",
      "GTEx bulk tissue TPM",
      "single-cell atlases"
    ],
    "graph_field": "data.cell_type.expressed_in"
  },
  {
    "class": 2,
    "name": "required_in",
    "strength": "moderate to strong",
    "typical_sources": [
      "DepMap gene effect split by lineage",
      "conditional knockout",
      "competitive transplantation"
    ],
    "graph_field": "data.cell_type.required_in"
  },
  {
    "class": 3,
    "name": "maintains_state",
    "strength": "strong when shown, usually absent",
    "typical_sources": [
      "inducible knockout with a state readout",
      "serial limiting-dilution transplant",
      "sorted-population profiling"
    ],
    "graph_field": "data.cell_state.maintains"
  }
]
two axes
{
  "cell_type": "what kind of cell - pro-B, GMP, HSC, LSC; node family cell-type-*",
  "cell_state": "what condition it is in - quiescent, differentiation-blocked, drug-tolerant, interferon-high; node family cell-state-*",
  "note": "orthogonal; a gene can be type-restricted and state-agnostic or the reverse, and collapsing them loses the distinction the PI asked for"
}
How this could bite us back
  • A target nominated on class-1 evidence alone will fail in the clinic without ever having been wrong on its own terms - the expression was real, the requirement was never tested.
  • A class-2 dependency that is pan-essential offers no therapeutic index; the same measurement that nominates it also disqualifies it, and only the floor-compression check separates the two readings.
Still open
  1. How many of this graph's existing gene-to-cell edges are class 1 presented as class 2? The audit has not been run.
  2. Is there any gene in the core set for which a class-3 claim has been demonstrated in human infant KMT2A::AFF1 material rather than inferred from a murine model?
Every source behind this entry (4)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.dataset

    Splitting the KMT2A-rearranged DepMap arm by lineage reclassifies CEBPA and SPI1 from apparent KMT2A-selective dependencies to myeloid-identity artefacts, because they move +0.15 and +0.92 respectively in the B-lineage half.

    What this evidence cannot show

    Five B-lineage and eight myeloid lines in the original analysis, so each half is small. The split is by lineage LABEL rather than by an expression signature.

    Evidence that argues against this (1)
    • AUTHORS' OWN STATED LIMITATION of the instrument this reclassification runs on. The Chronos authors state that CRISPR loss-of-function screens 'exhibit a number of biases and artifacts that can confound the results', and report that their own method was selected partly for exhibiting the lowest COPY NUMBER and SCREEN QUALITY bias among evaluated methods - which is a statement that both biases are present and material in this data type. A split into five B-lineage and eight myeloid lines, compared by lineage LABEL rather than by an expression signature, is exactly the regime in which screen-quality variation between a handful of lines can produce a +0.15 or +0.92 shift. CEBPA and SPI1 may well be myeloid-identity artefacts; the demonstration is under-powered for the certainty with which this concept entry uses it as its worked example. Dempster JM et al. 2021
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    In DepMap CRISPR gene effect over 1,258 lines fetched on 2026-09-20, WDR5, KAT8, RBBP5, BRD4, ELL and PAF1 are each depleted below -0.5 in at least 93.4% of all lines, while MLLT3 is depleted in 0.4%.

    What this evidence cannot show

    Cancer cell lines in culture, not primary cells. A -0.5 cut is conventional and not a biological threshold.

    Evidence that argues against this (1)
    • The three-class scheme understates a fourth problem it cannot fix: an expression measurement can be wrong about the PROTEIN as well as about the requirement. Internal-promoter-driven, YEATS-domain-devoid isoforms of MLLT1 and MLLT3 are expressed tissue-specifically, so a class-1 observation on those genes may be measuring a functionally different molecule. Dailhau S et al. 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    In mice, LEDGF/p75 is dispensable for haematopoiesis but essential for MLL-rearranged leukemogenesis.

    PMID 29084774DOI
    What this evidence cannot show

    Murine system. Brief rule 73: this is mouse haematopoiesis and murine MLL-rearranged leukaemia, not infant B-lineage KMT2A::AFF1 in human cells. Abstract read; full text not opened in this beat.

    Evidence that argues against this (1)
    • PUBLISHED CONFLICTING RESULT: 'dispensable' is true of one compartment in a conditional model and false of the gene. Sutherland et al. disrupted Psip1/Ledgf by gene trap and found that the MAJORITY of homozygous mice died PERINATALLY; survivors showed low fertility, absent epididymal fat pads, motor and behavioural defects including hind-limb clenching and reduced grip strength, brachycephaly, small rib cages and HOMEOTIC SKELETAL TRANSFORMATIONS - which the authors read as a role for Psip1 in the control of HOX EXPRESSION, and link explicitly to PSIP1/LEDGF appearing as a NUP98 fusion partner in myeloid leukaemia. Hox control is the same function the leukaemia argument invokes. So the cleanest example of a cell-context dissociation in this concept entry is cleanly conditional-model-specific, and the gene is not a normal-cell-dispensable gene. The graph also already carries `mechanism-psip1-in-normal-haematopoiesis-dispensable-is-not-silent`, which records halved white counts, perturbed HSC and progenitor numbers and failing serial replating WITHIN haematopoiesis - so this concept entry's exemplar of a clean class dissociation is the graph's own documented example of a summary that hides a phenotype. Sutherland HG et al. 2006
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    MLLT3 governs human haematopoietic stem-cell self-renewal and engraftment - a normal-cell function of the wild-type protein.

    PMID 31776511DOI
    What this evidence cannot show

    Human cord-blood and fetal HSCs with overexpression and knockdown; this is normal haematopoiesis, not leukaemia. Abstract read; full text not opened in this beat.

    Evidence that argues against this (1)
    • METHODOLOGICAL DEFEATER for assigning this result to evidence class 2 ('required') on the strength of the perturbation used. The human HSC result rests on overexpression and knockdown of MLLT3, and MLLT3/AF9 is a shared subunit of at least three distinct elongation complexes: Luo et al. isolated SEC, SEC-L2 and SEC-L3 all containing AF9/MLLT3 with P-TEFb and ENL, and showed they control DIFFERENT subsets of genes. Overexpressing or depleting a shared subunit therefore changes the stoichiometry of three complexes at once, and the self-renewal phenotype belongs to whichever combination shifted. This is the same subunit-versus-complex problem the entry's own class scheme should flag, and the class scheme currently has no slot for it. Luo Z et al. 2012
Connected to