Preferential formation of NUP98-KDM5A condensates at specific H3K4me3-rich loci drives leukemogenic gene expression

Reviewer #1 (Public review):

Leukemia-driving NUP98 oncofusion proteins form chromatin-associated biomolecular condensates in the nucleus, and these structures are important for oncogenic transformation. Most NUP98 fusions do not contain domains that mediate the recognition of specific DNA elements. Instead, they entail domains that are important for chromatin regulation. For instance, the NUP98::KDM5A fusion features a fusion of the NUP98 N-terminus with the third PHD domain of the histone demethylase KDM5A. As PHD domains are critical for the recognition of methylated histones without any sequence specificity, it is not clear what controls the condensation and chromatin binding of NUP98::KDM5A, leading to the induction of oncogenic transcriptional programs.

In this work, the authors use a combination of cellular and in vitro studies to show that biomolecular condensation of NUP98::KDM5A is dependent on H3K4me3 binding. Their model proposes that concentration-dependent chromatin-associated condensation of NUP98::KDM5A depends on local densities of H3K4me3 on chromatin and the levels of the fusion oncoprotein. In line with this, the analysis of gene expression data from NUP98::KDM5A-positive AML cells shows a positive correlation between differentially expressed genes and H3K4me3 levels.

This is an interesting manuscript that aims to dissect the molecular mechanisms underlying biomolecular condensation of the NUP98::KDM5A oncoprotein. The work is solid, and the results are well explained and presented in a logical order. However, the study suffers from several weaknesses that if addressed would improve the study.

Major points:

(1) All cellular experiments are performed in settings of transient transfection of NUP98::KDM5A in non-hematopoietic cell types. These conditions are not physiologically relevant, as these cells do not depend on the fusion oncogene. Therefore, any claims about concentration-dependent effects on condensation need to be validated in AML cells that are driven by NUP98::KDM5A. While this may not be possible in primary patient-derived cells, several groups have published AML models of NUP98::KDM5A-driven AML that could be used.

(2) The results presented in Figure 4 are not entirely supportive of the mechanism. It is known that active gene expression correlates with high H3K4me3 levels; therefore, the correlations shown by the authors are expected. Yet, the authors do not discuss the fact that many H3K4me3-positive genomic regions do not show NUP98::KDM5A binding. This should be elaborated on in the discussion section.

(3) While the focus of the manuscript is on NUP98::KDM5A, this oncofusion is part of a family of >30 fusions that join the NUP98 N-terminus to a variety of factors with roles in epigenetic control and transcription. While the repertoire of NUP98 fusion partners is diverse with regard to functional domains, they all induce a conserved set of target genes that is characteristic of this leukemia subtype. How can this be achieved in the context of NUP98 fusion proteins that do not contain a PHD domain, such as NUP98::NSD1 or NUP98::HOXA9? Please discuss this.

Reviewer #2 (Public review):

In this manuscript, the authors investigate how the oncogenic fusion protein NUP98-KDM5A alters gene expression in leukemia, using a combination of cellular experiments with model and patient cell lines, as well as in vitro studies. Upon transfection of U2OS cells with mEGFP-tagged NUP98-KDM5A, the authors show that the fusion proteins form sub-micrometer puncta, whereas KDM5A alone does not. These foci are also observed at expected native expression levels (using OpenCell data). The tag has an effect here, as switching to an mCherry tag raises the apparent saturation concentration for phase separation. Finally, the authors show via super-resolution imaging that the foci correlate with H3K4me3 distribution.

In vitro, the fusion protein forms amorphous, gel-like condensates at double-digit nanomolar concentrations. Truncation analysis identifies PHD3 of KDM5A as required for maximal phase separation, consistent with the ability of the protein to bind H3K4me3 peptides. Addition of polynucleosomes increases the amount of fusion protein partitioning into the condensate in an H3K4me3-binding-dependent manner. Condensates are gel-like with slow internal dynamics in vitro; in cells, however, the dynamics depend on the position of the EGFP tag (no corresponding experiments with mCherry are shown). Reconstitution with H3K4me3- and H3K4me0-modified arrays shows colocalization with both wild-type NUP98-KDM5A and the binding mutant. Here, H3K4me3 arrays recruit ~20% more protein and yield gel-like structures in a manner dependent on the PTM and on the PHD finger.

In cells, the fusion protein colocalizes with H3K4me3-marked loci, including the HOX clusters, as confirmed by FISH. Finally, re-analysis of published expression datasets from patient cells shows that genes are predominantly upregulated and that the upregulated genes are H3K4me3-marked.

This is a well-executed mechanistic study. The data convincingly establish that NUP98-KDM5A forms sub-micrometer foci at realistic expression levels, that these foci correlate with H3K4me3-marked sites, that the PHD3-H3K4me3 interaction mediates chromatin binding while the NUP98 moiety drives phase separation in vitro, that foci in cells overlap genes heavily decorated with H3K4me3, and that H3K4me3-marked genes are those found to be upregulated in patient datasets. These are important mechanistic findings and of interest to the community.

Still, the functional/causal link is a bit more tentative, as the data is mostly correlative, since it is not directly established that there is feedback between H3K4 methylation, NUP98-KDM5A recruitment, phase separation and target gene overexpression. An experiment that could further bolster this claim would be a direct test of whether NUP98-KDM5A expression drives overexpression of bound genes, e.g. expression of the fusion protein vs PHD- and NUP98-mutant variants, followed by qPCR of target genes, such as the HOX cluster, and possibly H3K4me3 ChIP at the same loci. As all the constructs and cell lines exist, this could be feasible and would substantially strengthen the manuscript.

Author response:

We greatly appreciate the positive and constructive comments from the reviewers, which recognized the intellectual contributions of our manuscript and also captured its limitations. Below is our response to the three major points from Reviewer #1 and the comments from Reviewer #2.

Response to Reviewer #1

(1) Use of transient transfection in non-hematopoietic cells. We appreciate the reviewer’s point regarding physiological relevance. Our goal in these cellular microscopy experiments was to dissect the biophysical principles of NUP98::KDM5A (including its mutants) condensate formation under controlled expression levels (concentration). While AML model systems driven by NUP98::KDM5A are available, they do not offer this possibility because of pre-existing NUP98::KDM5A expression. The suspension culture of hematopoietic cells also brings practical challenges for correlative FISH+IF and high-resolution microscopy analysis. We agree that validating our observed behaviors in AML models would be valuable, e.g. by creating HSPC lines with inducible expression of tagged NUP98::KDM5A and its mutants, but such experiments fall outside the scope of the current study. We will add text acknowledging this limitation and clarifying that our mechanistic conclusions are grounded in biophysical principles that should generalize across cell types.

(2) Correlation between H3K4me3 and gene activation. We agree that active transcription correlates with H3K4me3, and that this baseline relationship must be considered. Our analysis explicitly uses fold-change between patient cells and healthy controls as the readout. This comparison inherently accounts for the activating effect of H3K4me3 itself. Regarding the reviewer’s comment that “many H3K4me3-positive genomic regions do not show NUP98::KDM5A binding”, we would like to clarify that this point is exactly what our manuscript aims to explain. Our cell line studies demonstrate that, at a patient-relevant expression level, NUP98::KDM5A condensates form preferentially at H3K4me3 locus with high local mark density. Considering that NUP98::KDM5A concentration in the nucleus is lower than the K_D between KDM5A PHD3 and H3K4me3, this means that H3K4me3 loci with lower mark density will not see NUP98::KDM5A binding without the high local concentration of the fusion protein (as a result of condensate formation). This is consistent with our observation that genes with the highest local density show disproportionately stronger upregulation. We will further clarify this point in the revised manuscript.

(3) Generalization to other NUP98 fusions lacking PHD domains. We appreciate this important conceptual question. We will expand the discussion to note that many NUP98 fusions, despite diverse partner domains, produce similar transcriptional programs. Our current hypothesis is that the highly active status of the HOX cluster genes in HSPC attracts NUP98 oncofusions targeting H3K4me3 (e.g. KDM5A and PHF23), while NUP98::HOXA9 and other transcription factor fusions directly target the HOX cluster via DNA sequence recognition. This convergence and the broad targets of the dysregulated HOX transcription factors explain the transcriptional program similarity, sustained by the previously described enrichment of transcriptional co-activators by NUP98 oncofusion condensates. We will explicitly discuss this hypothesis in our revised manuscript.

Response to Reviewer #2

We thank the reviewer for the thoughtful evaluation and agree that the causal link between H3K4me3 recognition, condensate formation, and gene activation remains partly correlative. Experiments such as qPCR or ChIP following expression of WT versus mutant constructs would indeed strengthen the causal chain. However, performing these assays across multiple constructs and loci in a physiologically relevant system is not feasible within the current revision cycle. We have added text acknowledging this limitation and clarifying that our study focuses on establishing the biophysical mechanism of targeting, while functional consequences are inferred from patient datasets rather than new perturbation experiments.