Oncofusion

start / graph / hyp-paralogue-buffering-hides-dependencies

Hypothesismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q22-target-rewrite

The dependencies this project keeps failing to see are the buffered ones: paralogue redundancy and guide placement, not absence of biology

also called paralogue buffering hypothesis · why CRISPR misses KDM4C and BOD1L

When we test whether a leukemia needs a gene, we delete one gene at a time and see if the cells die. That test is blind in two specific ways: it misses genes that have near-identical backups, and it misses genes where the cutting tool happened to land in a part that does not matter. Two candidates this cycle are invisible for exactly those reasons.

The same thing, for a scientist

Genome-wide CRISPR gene-effect screening is the project's default first filter on candidate targets, and two of this cycle's candidates score at or near zero in it while independent evidence says they matter. KDM4C is +0.009 in SEM and +0.002 in MV4-11 yet the KDM4 inhibitor SD70 preferentially kills SEM, and Agger et al. needed a conditional triple knockout to expose the family requirement in normal HSC. BOD1L1 is -0.150 in MV4-11 yet a 618-sgRNA tiling screen in the same line finds a strong dependency confined to one N-terminal domain of a 3,051-residue protein. The hypothesis is that these are two instances of one systematic bias with a shared prediction: a gene's Chronos score under-reports its requirement in proportion to its paralogue coverage and to the length over which a standard library's guides are distributed. If true, the project's target list has been pruned by an instrument whose blind spots are enumerable, and the pruning can be audited cheaply rather than repeated.

Go deeper
The two observations that generate it

KDM4C. DepMap gene effect across four KMT2A::AFF1 lines: SEM +0.009, SEMK2 +0.006, RS4;11 -0.157, MV4-11 +0.002; mean -0.017 across 1,258 lines with 0.8% depleted 1. Against that, SD70 preferentially suppressed SEM and THP1 over non-MLL lines, an independent KDM4C shRNA abolished leukemogenesis of primary human MLL leukemia cells in vivo, and the recruitment domain maps to the MLL 5' moiety retained in every fusion 2. The reconciling fact is that Agger et al. required a conditional triple knockout of KDM4A/B/C to expose the family requirement in HSC 3 — direct evidence that single-gene KDM4 loss is buffered in haematopoietic cells, and SD70 is a family inhibitor.

BOD1L1. Mean Chronos -0.234 with 6.0% depleted, and the primary paper agrees it is called non-essential in all leukemia lines. Yet a 618-sgRNA tiling screen in iCas9-MV4-11 and iCas9-MOLM-13 finds a strong dependency on the N-terminal Shg1 domain and none on the C-terminal disordered regions or AT-hook, and the authors state the reason directly: perturbation effects depend on sgRNA position 4.

These look like two anecdotes. The hypothesis is that they are one mechanism seen twice — the screen measures the average consequence of cutting a gene, and both paralogue coverage and gene length dilute that average away from the consequence of removing the function.

Why this matters more than the two genes

This project prunes candidates on Chronos routinely, and did so three times this cycle: CCNT1 (-0.206, 8.4%), ELL2 (-0.064, 1.0%) and ASH2L (-0.333, 21.4%) were all struck at least partly on a low gene effect. Two of those three are short proteins with no close paralogue and the objection does not reach them. ELL2 is not one of them. ELL2 has a paralogue — ELL, at -1.302 with 94.0% of lines depleted — and the strike reasoning explicitly uses the contrast between the two as evidence that ELL2 is the dispensable one.

Under this hypothesis that inference is backwards in a specific and testable way. If ELL and ELL2 were mutually buffering, the expected pattern would be that neither scores; the observed pattern is that one scores strongly and the other not at all, which is what non-redundancy looks like, not redundancy. So the hypothesis, applied honestly, supports the ELL2 strike rather than undermining it — and that is the point of writing it down as a hypothesis rather than as a rhetorical device. It has to be able to come out the other way.

What it does undermine is the general practice. Three failure modes are now on record for the essentiality axis of ana-target-safety-triage-v2: underpowered constraint (DPY30, v1), phenotypes the assay cannot measure (ELL2's humoral defect), and the two in this entry. None was caught by the rule; each was caught by somebody reading one gene's genetics.

What would kill this

The hypothesis is cheap to falsify and the falsifiers are not symmetric with the confirmations, which is what makes it worth stating.

It dies if, across the graph's candidate set, Chronos gene effect shows no relationship to paralogue identity or to coding length once expression is controlled. It dies for KDM4C specifically if re-expression of KDM4A or KDM4B fails to rescue SD70 sensitivity in SEM — that would mean SD70's effect is KDM4C-specific after all, and the CRISPR null would then be a genuine contradiction requiring a different explanation, most likely that SD70 is off-target. It dies for BOD1L if a tiling screen in FLT3-wild-type KMT2A::AFF1 lines finds no Shg1-domain dependency, which would relocate the MV4-11 result to FLT3-ITD.

Note what does not confirm it: finding that some buffered gene is a real dependency. One example is an existence proof of the mechanism, which is already established, not evidence that it is prevalent enough to have distorted this project's target list. The claim that matters is the prevalence claim, and only the systematic analysis addresses it.

Facts and figures
statementGenome-wide single-gene CRISPR gene effect systematically under-reports a gene's requirement in proportion to (a) the coverage of close paralogues that can substitute for its function and (b) the coding length over which a standard library distributes few guides relative to the size of the functional element. Candidates pruned from this project's target list on a low Chronos score are therefore enriched for buffered true dependencies.
origindata-pattern
prior probability0.6
predictions
[
  "Across this graph's candidate set, Chronos gene effect correlates positively (less negative) with coding length after controlling for expression.",
  "Genes with a close paralogue expressed in the same lines show systematically weaker gene effect than genes without one, matched on complex membership.",
  "Re-expressing KDM4A or KDM4B in SEM rescues SD70 sensitivity; re-expressing a catalytically dead KDM4C does not.",
  "A CRISPR tiling screen of BOD1L in FLT3-wild-type KMT2A::AFF1 lines (SEM, RS4;11) reproduces the Shg1-domain dependency seen in MV4-11.",
  "The mechanism does NOT rescue ELL2: ELL and ELL2 show a split rather than a mutual null, which is the signature of non-redundancy."
]
experiments
[
  "exp-kdm4-paralogue-rescue-in-sem"
]
analyses
[
  "ana-target-list-rewrite-audit"
]
falsifiers
[
  "Across the graph's candidate set, Chronos shows no relationship to coding length or to paralogue presence once expression is controlled. The hypothesis is then an anecdote about two genes.",
  "Re-expression of KDM4A or KDM4B fails to rescue SD70 sensitivity in SEM. The KDM4C CRISPR null then genuinely contradicts the pharmacology and the likeliest remaining explanation is that SD70 is off-target — which would remove KDM4C from the list rather than rescue it.",
  "A BOD1L tiling screen in SEM and RS4;11 finds no Shg1-domain dependency, relocating the MV4-11 result to FLT3-ITD and removing the KMT2A::AFF1 claim entirely.",
  "Published re-analysis shows Chronos already corrects for guide-position and paralogue effects to a degree that makes the residual bias negligible."
]
How this could bite us back
  • The hypothesis is a licence to reopen strikes, and a licence to reopen anything is a licence to reopen everything. Without the prevalence analysis it can be used to argue that any inconvenient null is an artefact, which is the opposite of what a falsifiable claim should enable. It should not be cited to reinstate a candidate until the systematic analysis exists.
  • If buffering is why a target looks dispensable, it is also usually why hitting it will not work: a buffered node is a node the cell can route around, and a family inhibitor broad enough to overcome the buffering is broad enough to hit the normal compartment. KDM4C is the worked example — the redundancy that explains the CRISPR null is the same redundancy that makes the therapeutic window depend on the family knockout phenotype in HSC.
  • The two genes generating this hypothesis are the two genes this cycle created entries for, which is a conflict of interest in the reasoning. A mechanism discovered while justifying two new candidates needs testing against a set that was not selected for it, and E5 is the only such test performed here.
Still open
  1. Is there a published systematic estimate of paralogue buffering in DepMap that would make the prevalence analysis unnecessary? If one exists, this hypothesis should be replaced by a citation.
  2. How many of this graph's 435 entries contain a dispensability claim resting on a Chronos value, and how many of those genes have a close expressed paralogue? That count is a morning's work and nobody has it.
  3. Does the guide-placement bias have a direction that matters clinically? A target that a standard library misses is also a target whose functional domain is small and well defined, which is exactly what a small-molecule programme wants.
  4. If buffered dependencies are systematically invisible, the project's pruning has been biased toward targets that are singletons and short. Is the surviving list unrepresentative in a way that would be visible if anyone looked?
Every source behind this entry (5)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    KDM4C shows no DepMap dependency in any KMT2A::AFF1 line (SEM +0.009, SEMK2 +0.006, RS4;11 -0.157, MV4-11 +0.002) with a mean of -0.017 across 1,258 lines and 0.8% depleted; BOD1L1 shows a mean of -0.234 with 6.0% depleted and -0.150 in MV4-11.

    What this evidence cannot show

    Queried by this agent this cycle; DepMap release not stated by the API and therefore unverified. These are the observations the hypothesis exists to explain, so they cannot also be evidence for it.

    Evidence that argues against this (1)
    • These numbers are model output, not measurements, and the model exists because the raw screen was systematically wrong in a way that looks like biology. Meyers et al. built CERES after showing that a gene-INDEPENDENT antiproliferative effect of Cas9 cleavage confounds dependency measurement and produces false positives in copy-number-amplified regions; they screened 342 lines, applied the correction, and showed it reduced false positives and estimated sgRNA activity across libraries. A per-line gene effect near zero is therefore a residual after two corrections - for copy number and for guide activity - and the hypothesis's own claim is that the screen is blind in a third way the model does not correct for. Quoting Chronos values to three decimal places overstates what the instrument resolves. Meyers RM et al. 2017
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The KDM4C inhibitor SD70 preferentially suppressed human leukemia cell lines carrying MLL fusions, naming SEM and THP1, over non-MLL lines; an independent KDM4C shRNA in primary human MLL leukemia cells left the entire knockdown cohort alive through 90 days against a control cohort that died; and the KDM4C interaction domain maps by deletion to the MLL 5' moiety present in all MLL fusions.

    PMID 26766589DOI
    What this evidence cannot show

    Full text read (PMC4712026). The SEM result is a viability panel with no transcriptional readout and no paralogue rescue, which is exactly the experiment this hypothesis calls for. The 90-day survival result is in a primary MLL-fusion AML sample whose partner is not specified in the text, not in a KMT2A::AFF1 model. **Resolved:** E-MTAB-3322 is 16 of 16 Mus musculus, and by the sweep's reading of the sample files comprises MOZ-TIF2 (8), MLL-GAS7 (6) and MLL-AF9 (2) - so **half the series is not a KMT2A fusion**. The human 'MLL3' xenograft quoted alongside deposited nothing and its partner is unnamed even in full text. Resolved 2026-09-19 by the Q18b unstated-systems sweep from the study's deposition, not from its text; bounded as a statement about the deposited data rather than about the whole study.

    Evidence that argues against this (2)
    • The positive side of the contradiction is as fragile as the negative side, and this hypothesis needs both to be solid. Ben-David et al. showed that strains of the same cell line diverge genetically and transcriptionally under culture, and that across 321 anti-cancer compounds at least 75% of compounds strongly inhibiting some strains of one line were completely inactive in other strains of the same line. The SD70 preference for SEM and THP1 over non-MLL lines is a single-laboratory viability panel with no transcriptional readout; if it does not reproduce in independently sourced lines, the discrepancy this hypothesis explains evaporates and no buffering needs to be invoked. Testing that is cheaper than the paralogue rescue and should precede it. Ben-David U et al. 2018
    • A paralogue-buffering argument built on this series is averaging across three different fusion oncoproteins, one of which (MOZ-TIF2) does not involve KMT2A and one of which (MLL-GAS7) has documented independence from the Hoxa7/Hoxa9 axis the others require. Buffering that appears across such a mixture may be a property of the mixture. EMBL-EBI BioStudies 2026
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Exposing the KDM4 family requirement in haematopoietic stem cells required conditional TRIPLE-knockout mice, which is direct evidence that single-gene KDM4 loss is buffered in haematopoietic cells.

    The KDM4/JMJD2 histone demethylases are required for hematopoietic stem cell maintenance · Agger K, Nishimura K, Miyagi S, Messling JE, Rasmussen KD, Helin K · Blood 2019
    PMID 31434704DOI
    What this evidence cannot show

    Abstract read, not full text. The inference that single-gene loss is buffered is drawn from the authors' experimental design rather than from a stated single-versus-triple comparison in the abstract; the abstract does not report single-knockout controls, so 'they needed a triple knockout' is read from what they did, not from a reported negative result. Mouse, and normal haematopoiesis rather than leukemia.

    Evidence that argues against this (1)
    • 'They needed a triple knockout, therefore single-gene loss is buffered' does not follow, and the counterexample is in this graph. Chen et al. deleted Mll1 and its closest paralogue Mll2 in MLL-AF9-transformed cells: Mll2 deletion ALONE significantly impaired survival, additional Mll1 loss further reduced viability and proliferation, and the authors showed explicitly that the collaboration is not redundancy but regulation of distinct pathways. Two paralogues can therefore be jointly required without either substituting for the other, and a multi-gene knockout requirement can reflect additive contributions to different pathways rather than buffering. Distinguishing the two needs the single-knockout arms reported, which the KDM4 triple-knockout abstract does not give - so the inference here rests on experimental design, not on a measured negative. Chen Y et al. 2017
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    A 618-sgRNA CRISPR tiling screen against BOD1L in doxycycline-inducible Cas9 MOLM-13 and MV4-11 leukemia cells identified a strong dependency on the N-terminal Shg1 domain but not on the C-terminal intrinsically disordered regions or AT-hook, with the authors noting that CRISPR perturbation effects depend on sgRNA position and that DepMap classifies the gene as non-essential in all leukemia cell lines.

    BOD1L mediates chromatin binding and non-canonical function of H3K4 methyltransferase SETD1A · See PubMed record PMID 38989615 for the full author list (not transcribed here) · Nucleic Acids Research 2024
    PMID 38989615DOI
    What this evidence cannot show

    Full text read (PMC11381347). A tiling screen is a within-gene comparison and is not immune to the artefacts of any CRISPR screen; a region of apparent dependency can also reflect local cutting toxicity or a copy-number effect, neither of which the abstract addresses. Both screened lines carry FLT3-ITD.

    Evidence that argues against this (1)
    • The competing explanation for a domain-localised tiling signal on a very long protein is not biology but assay artefact, and the paper's own design partly controls for it: the Shg1-domain guides' effect was reproduced by selected targeting sgRNAs that reduced BOD1L protein, and BOD1L knockout phenocopied SETD1A knockout transcriptionally. A pure cutting artefact would not produce a matching transcriptional signature. That is reassuring but not decisive, and it comes from the same paper rather than from an independent group. Nucleic Acids Research 2024
  5. E5moderate confidence 3 of 5: One strong study, or several smaller consistent ones.database

    ELL2 is the case where this hypothesis predicts the opposite of what it might be used to argue: ELL and ELL2 are paralogues, yet ELL scores -1.302 with 94.0% of lines depleted while ELL2 scores -0.064 with 1.0% — a split, not the mutual null that reciprocal buffering would produce.

    What this evidence cannot show

    Values taken from an existing entry in this graph rather than re-queried this cycle. The argument is inferential: reciprocal buffering between two paralogues predicts that neither scores, but a one-way dependency (ELL needs no help, ELL2 cannot substitute) is also consistent with a split, and expression differences across lines could produce the same pattern without any statement about redundancy. This is the weakest evidence item here and is included because the hypothesis has to be applied to the case where it argues against the writer's convenience.

    Evidence that argues against this (1)
    • The ELL/ELL2 split may not be a biological split at all. Meyers et al. showed that CRISPR essentiality screens carry a gene-independent antiproliferative signal from Cas9 cleavage that scales with copy number and produces false positives in amplified regions, and that sgRNA activity varies enough between libraries to need explicit estimation; CERES corrects both, imperfectly. A gene scoring -1.302 with 94% of lines depleted against a paralogue at -0.064 is exactly the shape that differential guide efficacy or a copy-number residual would produce, and this item already concedes that expression differences would produce it too. Using it as the hypothesis's self-imposed hard case is honest, but it is not yet evidence in either direction. Meyers RM et al. 2017
Connected to