Condensates in KMT2A-fusion leukemia — an honest verdict: not established for AFF1/AFF4/AF9, load-bearing only in ENL-mutant and NUP98 fusions
also called phase separation MLL fusion · AFF1 condensate · AFF4 LLPS · ENL condensate · NUP98 phase separation
There is an exciting idea that fusion proteins work by forming liquid droplets in the nucleus. For KMT2A fusions this has essentially not been shown. The good droplet evidence comes from different proteins — mutant ENL and NUP98 fusions — and should not be borrowed.
The same thing, for a scientist
Direct answer to the question as posed: phase separation is NOT established as mechanistically load-bearing for KMT2A fusions, and for AFF1/AF4, AFF4 and AF9/MLLT3 it is not established at all. A dedicated verification pass returned zero PubMed-indexed papers reporting AFF1, AFF4 or AF9 condensates in leukemia, and no KMT2A-fusion condensate separation-of-function mutant exists. Load-bearing condensate genetics does exist, but for two other oncoprotein classes: ENL YEATS hotspot point mutants (in vivo knock-in, condensate-disrupting mutagenesis impairs oncogenic function, an acetyl-binding inhibitor displaces condensates) and NUP98 fusions (FG-repeat IDRs essential for LLPS and transformation, rescued by an unrelated FUS IDR). Even there, one verified study shows condensate formation is necessary but not sufficient.
The negative result, stated precisely because it matters1 deeper
A dedicated verification worker searched PubMed for AFF1/AF4 phase separation, MLL-AF4 condensates, AFF4 phase separation and AF9/MLLT3 phase separation. The searches for AFF1 phase separation, MLL-AF4 condensate/phase separation and AF9/MLLT3 phase separation returned zero results. The only AFF4-adjacent hit was a study of FUS/AFF4 co-localisation in nuclear puncta in HIV latency, disrupted by aliphatic alcohol (i.e. 1,6-hexanediol-class evidence, the weakest tier) — not leukemia, not KMT2A 1.
So any node, review or grant paragraph asserting that AFF1/AFF4 condensates drive KMT2A-fusion leukemia is unsupported by the indexed literature as of this cycle. This Queen recommends the graph record this explicitly as a negative, because the claim circulates widely by analogy and the analogy crosses oncoprotein classes with different architecture.
The architectural point is worth making: AFF1 has a large intrinsically disordered region and a pSER domain, so it is a plausible condensate former, and the reciprocal AFF1::KMT2A product is also IDR-rich. Plausibility is not evidence. Nobody has looked properly.
The methodological standard this field should be held to
The single most useful methodological finding comes from the ENL work and applies to every condensate claim: expression of ENL mutants beyond the endogenous level leads to non-functional condensates 2. The authors state it themselves.
That sentence invalidates a large fraction of the puncta literature. Ranking evidence tiers, from weakest to strongest:
- Weakest: in-vitro droplets of purified protein; 1,6-hexanediol or aliphatic-alcohol sensitivity; GFP puncta in overexpressing cells.
- Middle: endogenous-level tagging showing puncta that co-localise with function.
- Strongest: separation-of-function mutants that abolish condensate formation while preserving binding (or vice versa), tested at endogenous levels, with an in-vivo phenotype; plus IDR-swap rescue with an unrelated IDR.
Only tier 3 supports a mechanistic claim, and in KMT2A-fusion biology there is nothing at tier 3, tier 2, or arguably tier 1.
Where condensate biology IS load-bearing — and why it should not simply be transferred
ENL YEATS hotspot mutants (Wilms tumour / AML, not KMT2A fusions). ENL YEATS hotspot mutations trigger aberrant condensates at native genomic targets via multivalent homotypic and heterotypic interactions; the required elements are the mutation-induced YEATS structural change, ENL's two disordered regions of opposing charges, and incorporation of extrinsic elongation factors; and extensive mutagenesis establishes condensate formation as a driver of oncogenic gene activation 2. A conditional knock-in mouse then showed mutant ENL forms condensates in HSPCs at leukemogenic loci, disrupting condensate formation by mutagenesis impairs chromatin and oncogenic function, and an acetyl-binding inhibitor displaces the condensates from target loci and delays AML in vivo 3. A further study showed mutant ENL binds RNA through a basic patch in the YEATS domain, blocking RNA binding impairs condensate reformation at endogenous targets, and disrupting RNA binding suppresses leukemogenesis in mouse models 4.
That is tier-3 evidence — but for a point-mutant chromatin reader, not for a chimeric fusion protein.
NUP98 fusions — the clean comparator. The NUP98 FG-repeat IDRs in NUP98-HOXA9 are essential for establishing LLPS puncta and for inducing leukemic transformation; LLPS promotes chromatin occupancy and is required for the broad super-enhancer-like binding pattern; and — the decisive control — replacing the NUP98 FG repeats with the unrelated LLPS-forming FUS IDR recapitulated enhanced genome-wide binding and target activation. Hi-C showed phase-separated NUP98-HOXA9 induces CTCF-independent chromatin loops enriched at proto-oncogenes 5. Independently, both homotypic and heterotypic interactions are required for NUP98-HOXA9 nuclear puncta, and three further NUP98 fusions also form puncta and transform, indicating LLPS is critical for leukemogenesis by NUP98 fusion oncoproteins 6.
And even there, necessary is not sufficient. In NUP98-NSD1, interactome binding is largely dependent on the FG repeats that mediate LLPS, and SMARCA5 is a functionally essential condensate constituent — but SMARCA5 inhibition did not affect the condensates, leading the authors to conclude that functional activity rather than condensate formation per se is crucial, and that condensate formation is not sufficient for maintaining the transformed phenotype 7.
The conceptual lesson that transfers, even though the data do not: NUP98 fusions carry their own IDR and need it; KMT2A fusions carry a chromatin-targeting head plus an elongation-machinery tail and are fully explained, at present, by recruitment biochemistry. Different architecture, different mechanism class. A borrowed condensate model risks redirecting effort away from the interface work that is actually supported.
Separation of function: compartment address is not transformation (added 2026-09-18)
Counter-evidence added 2026-09-18: the compartment address and the transforming function are separable. Ayton, Chen & Cleary mutagenised the retained MLL N-terminus and found three regions, including CXXC/Basic, that are essential for myeloid transformation by mechanisms that do not compromise subnuclear localisation; CXXC point mutants that kill transformation also kill nonmethylated-CpG binding and transactivation 8.
This is the cleanest separation-of-function result available on this question, and it cuts a specific way. A fusion can hold its subnuclear position and still be transformation-dead. So being in the right compartment is not sufficient, and any hypothesis that makes compartment occupancy - condensate, speckle, body - the driver of transformation has to explain why the address can be kept while the function is lost. What the same experiment does not do is exonerate localisation generally: the CXXC mutants lose the locus address (CpG binding) at the same time. The result separates compartment from transformation; it leaves locus and transformation firmly coupled.
| chain | mechanism-kmt2a-fusion-causal-chain |
|---|---|
| verdict for kmt2a fusions | not established; for AFF1, AFF4 and AF9/MLLT3 in leukemia there are zero indexed reports |
| verdict confidence | this is an absence-of-evidence finding from a systematic PubMed pass in cycle 2, not proof of absence |
| necessary but not sufficient | NUP98-NSD1: SMARCA5 inhibition abolishes function without affecting condensates |
| methodological standard | supra-endogenous expression produces non-functional condensates (authors' own finding), so overexpressed-GFP-puncta and 1,6-hexanediol evidence is not admissible |
| kmt2a fusion evidence tier available | none |
| architectural note | NUP98 fusions carry their own IDR; KMT2A fusions carry a targeting head plus an elongation-machinery tail and are currently fully explained by recruitment biochemistry |
load bearing elsewhere
[
{
"class": "ENL YEATS hotspot point mutants",
"tier": 3,
"evidence": "in vivo knock-in, condensate-disrupting mutagenesis, RNA-binding separation of function, inhibitor displaces condensates"
},
{
"class": "NUP98 fusions",
"tier": 3,
"evidence": "FG-repeat IDR essential; unrelated FUS IDR rescues; CTCF-independent loops"
}
]evidence tiers
{
"tier1_weak": "in vitro droplets; 1,6-hexanediol; overexpressed GFP puncta",
"tier2_middle": "endogenous-level tagging with functional co-localisation",
"tier3_strong": "separation-of-function mutants at endogenous levels with in vivo phenotype; IDR-swap rescue"
}- Adopting a condensate model for KMT2A fusions on the strength of NUP98 and ENL-mutant data would redirect effort away from the interface biology that is actually supported — a strategic, not experimental, risk.
- Condensate-dissolving agents are intrinsically non-selective: the same physical chemistry underlies nucleoli, speckles, stress granules and Pol II clusters, so on-target toxicity is likely if any such agent were pursued.
- Even where condensates are load-bearing, they can be necessary without being sufficient (NUP98-NSD1/SMARCA5), so dissolving a condensate may not disable the function it hosts.
- Every overexpression-based puncta result in this field is suspect because supra-endogenous expression produces non-functional condensates; a positive KMT2A-fusion result obtained by overexpression should not be believed.
- Does KMT2A::AFF1 form condensates at endogenous levels at its own target loci? Nobody has run the experiment. An endogenously tagged knock-in in a t(4;11) line with live imaging and an AFF1 IDR-deletion separation-of-function mutant would answer it definitively, and the answer is informative either way.
- Is the reciprocal AFF1::KMT2A product, which is IDR-rich, the condensate-competent partner rather than the KMT2A::AFF1 product? This has never been asked and would connect to mechanism-reciprocal-kmt2a-fusion-chromatin-opening already in the graph.
- If KMT2A fusions do not use condensates, why not, given that AFF1 has a large IDR? A negative answer would be a genuine mechanistic distinction between fusion classes and worth publishing.
- Does the ENL acetyl-binding inhibitor class (TDI-11055, SR-0813) work in KMT2A-r leukemia through condensate displacement or simply through reader-domain blockade? The two mechanisms predict different resistance profiles.
- Half-bleach internal mixing is the only assay shown to discriminate phase separation from clustered binding. Has it ever been applied to a KMT2A fusion, a SEC component, or NUP98-FG puncta? If not, every condensate claim in this graph rests on assays that cannot tell the two apart.
- E1weak confidence 2 of 5: Preliminary, small, indirect, or not yet peer reviewed.paper
FUS and AFF4 co-localise in nuclear puncta in the context of HIV-1 latency, with the puncta disrupted by aliphatic alcohol — the only AFF4-adjacent condensate report identified, and one that is not in leukemia and does not involve a KMT2A fusion.
Fused in sarcoma silences HIV gene transcription and maintains viral latency through suppressing AFF4 gene activation · Krasnopolsky S, Marom L, Victor RA, et al. (Taube R, senior author) · Retrovirology 16(1):16 2019What this evidence cannot show
Cited here to document what the only available AFF4 puncta evidence actually is, so that it is not repurposed. HIV latency in transformed cell lines; aliphatic-alcohol sensitivity is the weakest class of condensate evidence and is now widely regarded as non-specific. No leukemia, no KMT2A fusion, no separation-of-function genetics.
Evidence that argues against this (1)
- This is an HIV LATENCY study in a retrovirology journal, and the entry rightly flags it as the only FUS/AFF4 puncta evidence - but the disqualifying detail is the probe. Puncta disrupted by aliphatic alcohol means 1,6-hexanediol, which is not a specific condensate test: it perturbs the nuclear pore, inhibits kinases and phosphatases and disrupts chromatin organisation. Disruption by hexanediol is close to uninformative about phase separation. STRENGTHENED 2026-09-20: the objection above was that hexanediol is non-specific, and the benchmark literature is harsher than that. 1,6-hexanediol rapidly IMMOBILISES and CONDENSES chromatin in living human cells, an action the authors state is distinct from melting a droplet and which they instruct readers to use in reinterpreting chromatin-associated droplet results — so in a chromatin context the reagent does something close to the opposite of what the assay assumes. Life Science Alliance 2021
- E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
ENL YEATS hotspot mutations trigger aberrant condensates at native genomic targets via multivalent homotypic and heterotypic interactions requiring the mutation-induced YEATS structural change, ENL's two disordered regions of opposing charges, and incorporation of extrinsic elongation factors; extensive mutagenesis establishes condensate formation as a driver of oncogenic gene activation, and expression of ENL mutants beyond the endogenous level produces non-functional condensates.
Hotspot mutations in the structured ENL YEATS domain link aberrant transcriptional condensates and cancer · Song L, Yao X, Li H, et al. (Wan L, senior author) · Molecular Cell 82(21):4080-4098.e12 2022What this evidence cannot show
This concerns ENL YEATS point mutants in Wilms tumour and AML, NOT KMT2A fusion proteins, and must not be cited as KMT2A-fusion condensate evidence. The authors' own supra-endogenous-expression caveat is the standing methodological warning for the whole field.
Evidence that argues against this (1)
- These are HOTSPOT MUTANT ENL proteins. The condensate behaviour is a gain-of-function of the mutation, and KMT2A::ENL fusions carry wild-type ENL sequence, so this cannot be transferred to the fusion without assuming the mutation merely exaggerates a normal property. The paper's own framing - mutations LINK aberrant condensates to leukemogenesis - is about the mutants. Molecular Cell 2022
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
In a conditional knock-in mouse, mutant ENL perturbs haematopoiesis, expands myeloid progenitors and triggers rapid aggressive AML; mutant ENL forms condensates in HSPCs at key leukemogenic loci, disrupting condensate formation via mutagenesis impairs its chromatin and oncogenic function, and an acetyl-binding inhibitor displaces the condensates from target loci and delays AML initiation and progression in vivo.
Condensate-Promoting ENL Mutation Drives Tumorigenesis In Vivo Through Dynamic Regulation of Histone Modifications and Gene Expression · Liu Y, Li Q, Song L, et al. (Wan L, senior author) · Cancer Discovery 14(8):1522-1546 2024What this evidence cannot show
ENL point mutant, not a KMT2A fusion. The closest thing in the field to condensate formation being genetically load-bearing in vivo at endogenous expression — which is precisely why its absence for KMT2A fusions is notable rather than excusable.
Evidence that argues against this (1)
- The in vivo model is the same mutant ENL, not a KMT2A fusion. It establishes that a condensate-promoting ENL mutation is leukemogenic in mice, which is a strong result for that allele and silent on whether condensates are load-bearing for KMT2A::ENL - the question this entry exists to adjudicate. Cancer Discovery 2024
- E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Mutant ENL binds RNA via a basic patch in the YEATS domain; a chemically inducible condensate displacement and re-nucleation system shows that blocking ENL-RNA interaction or transcription impairs condensate reformation at endogenous targets, RNA binding preferentially raises mutant-ENL occupancy and transcriptional bursting at condensate-permissive loci, and disrupting RNA binding suppresses leukemogenesis in mouse models.
RNA reinforces condensate nucleation on chromatin to amplify oncogenic transcription · Budinich KA, Yao X, Gong C, et al. (Wan L, senior author) · Molecular Cell 86(14):2777-2793.e12 2026What this evidence cannot show
ENL point mutant again, not a KMT2A fusion. Very recent (2026), so independent replication does not yet exist.
Evidence that argues against this (1)
- Mutant ENL again, and the inducible displacement system is the entry's best tool precisely because it is not hexanediol - but it was applied to the mutant. Running the same displacement and re-nucleation on a KMT2A fusion is the experiment that would settle this entry's question and has not been done. Molecular Cell 2026
- E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The NUP98 FG-repeat intrinsically disordered regions in NUP98-HOXA9 are essential for establishing LLPS puncta and for inducing leukemic transformation; LLPS promotes chromatin occupancy and is required for the broad super-enhancer-like binding pattern; replacing the FG repeats with the unrelated LLPS-forming FUS IDR recapitulates enhanced genome-wide binding and target activation; and deeply sequenced Hi-C shows phase-separated NUP98-HOXA9 induces CTCF-independent chromatin loops enriched at proto-oncogenes.
Phase separation drives aberrant chromatin looping and cancer development · Ahn JH, Davis ES, Daugird TA, et al. (Wang GG, senior author) · Nature 595(7868):591-595 2021What this evidence cannot show
NUP98-HOXA9, an architecturally different oncoprotein class from KMT2A fusions. The IDR-swap rescue is the strongest separation-of-function design published in this space and is the standard a KMT2A-fusion condensate claim would have to meet.
Evidence that argues against this (1)
- In NUP98-NSD1, SMARCA5 inhibition did not affect the condensates, leading the authors to conclude that functional activity rather than condensate formation per se maintains the transformed phenotype, and that condensate formation is not sufficient. Jevtic Z et al. 2022
- E6strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Both homotypic and heterotypic interactions are required for NUP98-HOXA9 nuclear puncta, which are associated with aberrant transcriptional activity and transformation of HSPCs, and three further NUP98 fusion oncoproteins (NUP98-PRRX1, -KDM5A, -LNP1) also form nuclear puncta and transform, indicating LLPS is critical for leukemogenesis by NUP98 fusion oncoproteins.
Phase Separation Mediates NUP98 Fusion Oncoprotein Leukemic Transformation · Chandra B, Michmerhuizen NL, Shirnekhi HK, et al. (Kriwacki RW, senior author) · Cancer Discovery 12(4):1152-1169 2022What this evidence cannot show
NUP98 fusion class only. Puncta-to-transformation correlation across fusions is strong but the causal link rests on interaction mutants that may also affect binding.
Evidence that argues against this (1)
- A different fusion family, and NUP98's FG repeats are the special case rather than the general one - they are a dedicated low-complexity module that this graph's own partner work shows has no counterpart in the septin, ABI1 or gephyrin classes. Requirement of homotypic and heterotypic interactions for NUP98-HOXA9 puncta generalises to KMT2A fusions only if the partner supplies something comparable, and most do not. Cancer Discovery 2022
- E7moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
NUP98-NSD1 interactome binding is largely dependent on the NUP98 FG repeats that mediate LLPS condensates, and SMARCA5 and BPTF are condensate constituents with SMARCA5 functionally essential; however SMARCA5 inhibition did not affect the condensates, so functional activity rather than condensate formation per se is crucial, and condensate formation is not sufficient for maintenance of the transformed phenotype.
SMARCA5 interacts with NUP98-NSD1 oncofusion protein and sustains hematopoietic cells transformation. · Jevtic Z, Matafora V, Casagrande F, et al. (Schwaller J, Bachi A, senior authors) · Journal of Experimental & Clinical Cancer Research 41(1):34 2022What this evidence cannot show
Single study; the inference that condensates are insufficient rests on one pharmacological dissociation (SMARCA5 inhibition) rather than on a condensate-specific separation-of-function mutant. Still the best available necessary-but-not-sufficient citation in the field.
Evidence that argues against this (1)
- FG-dependence is not LLPS-dependence, and this is the cleanest confound in the entry. FG repeats mediate ordinary nucleoporin-transport-factor interactions as well as phase behaviour, so deleting them removes both. An interactome that collapses on FG deletion is consistent with condensates and equally consistent with loss of a conventional protein-protein interaction surface. Journal of Experimental & Clinical Cancer Research 2022
- E8strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
High-resolution mutagenesis of the retained MLL N-terminus identified three domains - the CXXC/Basic domain and two domains of unknown function - that are essential for in vitro myeloid transformation **via mechanisms that do not compromise subnuclear localization**; CXXC point mutants that abolish transformation also abolish binding to nonmethylated CpG DNA and transactivation in vivo.
Binding to nonmethylated CpG DNA is essential for target recognition, transactivation, and myeloid transformation by an MLL oncoprotein · Ayton PM, Chen EH, Cleary ML · Molecular and Cellular Biology 24(23):10470-8 2004What this evidence cannot show
ABSTRACT ONLY - the PMC full text was not read, so the specific microscopy controls behind 'do not compromise subnuclear localization' were not inspected; the imaging in that era was fluorescence microscopy of transfected constructs in COS cells, which is a coarse readout of compartment and cannot exclude changes in condensate number, size or dynamics. In-vitro myeloid immortalisation of mouse progenitors with an MLL fusion construct. The result separates gross subnuclear position from transformation; it does NOT separate chromatin occupancy from transformation - on the contrary, CXXC mutants lose CpG binding, so locus-level address and transformation stay coupled.
Evidence that argues against this (1)
- This mutagenesis identifies CXXC/Basic and two domains of unknown function as required for transformation - a result about DNA targeting, with no condensate readout. Used in a condensate entry it can only serve as an alternative explanation for the same phenotype, which is how the entry should file it. Molecular and Cellular Biology 2004
- E_benchmark_collapsestrong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
The assays this entry relies on to establish phase separation cannot distinguish it from the alternative. Model-free calibrated half-bleaching shows that BOTH 1,6-hexanediol sensitivity AND classical FRAP recovery fail to separate liquid-liquid phase separation from binding to spatially clustered sites; only half-bleach internal mixing discriminates the two. Independently, at physiological expression levels low-complexity-domain interactions that control transcription form dynamic hubs WITHOUT detectable phase separation.
Detecting and quantifying liquid-liquid phase separation in living cells by model-free calibrated half-bleaching · Nature Communications 2022What this evidence cannot show
Added by the coordinator on 2026-09-20 after a sweep surfaced the benchmark. This is a methodological result about assay discrimination, not a claim that any specific condensate in this graph is absent. It applies simultaneously to the SEC, PRC1/CBX2, NUP98-FG and ENL-hotspot claims held here, because all of them rest on hexanediol sensitivity, FRAP recovery or punctate appearance.
Evidence that argues against this (2)
- A stricter benchmark does not make the alternative correct. 'Binding to spatially clustered sites' is itself a substantive model and in several of the systems here it is arguably the more interesting one, since clustered binding predicts sequence-specific targeting where phase separation predicts bulk enrichment. The literature is not choosing between a real phenomenon and an artefact but between two mechanisms, and this entry should not read the benchmark as deflationary. Science 2018
- Dynamic hubs forming WITHOUT detectable phase separation was shown at physiological levels for specific low-complexity domains, and the fusion oncoproteins here are frequently expressed above physiological levels in the very systems used to study them. A result that depends on expression level cuts against overexpression studies and does not by itself settle what the endogenous protein does. Science 2018
- ← related toNUP98::KDM5A (NUP98-JARID1A), t(11;12)(p15;p13)
Condensate-standard bound (McSwiggen 2019, PMID 31594803): in-vivo LLPS evidence is often phenomenological and cannot discriminate phase separation from other mechanisms without endogenous-abundance measurement, and this claim rests entirely on ectopic expression.
NPM1c forms nuclear condensates, and the entry's hexanediol/FRAP caveat applies to that claim.
an honest verdict that the condensate case is not established here
The LLPS-sceptic position (McSwiggen 2019) is contested by comparable live-cell imaging of endogenous Mediator and Pol II condensates (Cho 2018, PMID 29930094); the field's consensus methods paper (Alberti 2019) asks for better controls, not rejection
supplies the methodological grounds for the not-established verdict
AKAP95 supplies partner-mediated condensate genetics the earlier verdict predates
a worked case where the condensate reading is correlative and the within-system AFF4 fluidity result is the load-bearing part
modifier, not a step - contested mechanism; not established for KMT2A fusions