NUP98 FG condensates, audited assay by assay: no half-bleach has ever been done and no NUP98 fusion has ever been endogenously tagged

The column that is empty across every row is half-bleach. That is the assay that discriminates 5, and in eight years of NUP98 condensate literature nobody has run it.

The purified-protein evidence is real and should not be discarded with the in-cell evidence.

• The Nature study's valency series is the strongest single argument in the field: purified constructs with 38 and 36 FG repeats formed droplets concentration-dependently while 27 and 11 FG repeats failed 2. A threshold in valency is what phase separation predicts and what simple aggregation does not. It is reinforced by a clean separation-of-function — an F→S mutation abolishing droplets without affecting stability, carried through ChIP-seq, RNA-seq and Hi-C — and by a FUS-IDR swap that phenocopies.
• The Cancer Discovery study quantified saturation concentration across FG mutants with recombinant protein: wild-type-like at under 10 nM, a 21×FG→AA mutant only above 160 nM, and a double mutant showing no condensation up to 20 µM 3, paralleling the in-cell constructs.

So: in a tube, with purified protein, the FG region of a NUP98 fusion undergoes concentration- and valency-dependent condensation. That is well supported. What is not supported is the step from there to "the endogenous fusion forms a liquid condensate in a leukaemia cell and that liquidity is the mechanism". The third paper's own observation that its condensates gel within minutes and do not fuse 3 argues directly against liquidity even in vitro.

Three of the four studies use 1,6-hexanediol as their principal in-cell test 124. In a chromatin context the reagent rapidly immobilises and condenses chromatin in living human cells, an action the authors describe as distinct from melting a droplet and which they explicitly instruct readers to use when reinterpreting chromatin-associated droplet results 6.

NUP98 fusion puncta are chromatin-associated by construction — they sit on Hox clusters. So hexanediol is acting on the very substrate whose clustering is the rival explanation. A punctum that disperses on hexanediol treatment is consistent with a melted droplet and with a chromatin template that has been reorganised out from under a clustered binder. The assay cannot separate them, and here it is being applied in the one context where the confound is maximal.

One methodological credit is due: the Molecular Cell study pairs hexanediol with spike-in-normalised ChIP-Rx and reports correlation of r = 0.89–0.94 with the FG→FA mutant chromatin profile 4. That is a more quantitative use of the reagent than eyeballing puncta, and it does show that hexanediol and the FG mutation produce the same chromatin outcome. It still does not distinguish the two models, because both predict that outcome.

The rival model is clustered binding to a template: the fusion binds many adjacent sites, and the resulting local concentration produces a punctum without a phase boundary.

In the NUP98 system this is not a deflationary alternative, it is a mechanistic proposal with more explanatory reach than phase separation has:

• Targeting. NUP98 fusions occupy specific loci — Hoxa, Hoxb, Meis1 — through partner-supplied readers and writers (the KDM5A PHD finger reading H3K4me3, the NSD1 SET domain). Clustered binding predicts sequence- and mark-specific targeting; bulk phase separation does not, and the partner's targeting module is the best-established part of NUP98 fusion biology.
• MLL1 recruitment. The competing accounts of MLL1 association (mechanism-nup98-fusion-mll1-nsl-interaction) look different under the two models: under clustered binding, MLL1 is recruited to specific sites; under phase separation, MLL1 partitions into a bulk phase along with every other coactivator. The Molecular Cell study's own framing — selective partitioning of transcriptional coactivators — sits closer to the phase model, while the ChIP-seq co-occupancy at 37 promoters sits closer to the clustered-binding one.
• Drug design. The two models give different targets. Phase separation nominates the FG region's bulk properties; clustered binding nominates the partner's reader domain and the sites it reads. Disulfiram directly targeting the PHF23 PHD finger is a clustered-binding-style intervention that already exists.

The experiment that decides it is available and cheap by the standards of this field: calibrated half-bleaching of an endogenously knock-in-tagged NUP98 fusion. The adjacent NPM1c literature shows what that looks like — CRISPR knock-in tagging, dTAG degron degradation, radial distribution functions and CUT&RUN, using neither FRAP nor hexanediol 8.

It does not say NUP98 fusion condensates are artefacts. It does not say the FG repeats are unimportant — the valency series, the F→S separation-of-function and the FG→AA saturation-concentration series all say they are essential. It does not say the published papers are wrong.

It says: the published in-cell evidence cannot distinguish the two models, the discriminating assay has never been run, and every in-cell claim rests on overexpression. Any entry in this graph that asserts NUP98 fusions act by liquid-liquid phase separation is asserting more than the evidence supports and should say 'condensate' or 'clustered occupancy' instead.