NSD2 inhibitors rewire chromatin to treat lung and pancreatic cancers

Abstract

NSD2 catalyses the epigenetic modification H3K36me2 (refs. 1,2) and is a candidate convergent downstream effector of oncogenic signalling in diverse malignancies^3–5. However, it remains unclear whether the enzymatic activity of NSD2 is therapeutically targetable. Here we characterize a series of clinical-grade small-molecule catalytic NSD2 inhibitors (NSD2i) and show that the pharmacological targeting of NSD2 constitutes an epigenetic dependency with broad therapeutic efficacy in KRAS-driven preclinical cancer models. NSD2i inhibits NSD2 with single-digit nanomolar half-maximal inhibitory concentration potency and high selectivity over related methyltransferases. Structural analyses reveal that the specificity of NSD2i for NSD2 is due to competitive binding with S-adenosylmethionine and catalytic disruption through a binary-channel obstruction mechanism. Proteo-epigenomic and single-cell strategies in pancreatic and lung cancer models support a mechanism in which sustained NSD2i exposure reverses pathological H3K36me2-driven chromatin plasticity, re-establishing silencing at H3K27me3-legacy loci to curtail oncogenic gene expression programs. Accordingly, NSD2i impairs the viability of pancreatic and lung cancer cells and the growth of patient-derived xenograft tumours. Furthermore, NSD2i, which is well-tolerated in vivo, prolongs survival in advanced-stage autochthonous KRAS^G12C-driven pancreatic and lung tumours in mouse models to a comparable level as KRAS inhibition with sotorasib⁶. In these models, treatment with both a NSD2 inhibitor and sotorasib synergize to confer sustained survival with extensive tumour regression and elimination. Together, our work uncovers targeting of the NSD2–H3K36me2 axis as an actionable vulnerability in difficult to treat cancers and provides support for the evaluation of NSD2 and KRAS inhibitor combination therapies in a clinical setting.

NSD2 (also known as MMSET and WHSC1) generates H3K36me2, an abundant histone modification that antagonizes EZH2-catalysed H3K27 methylation^7–13 (Extended Data Fig. 1a) and regulates DNA-templated processes, particularly transcriptional activation^14–16. Deregulation of NSD2 through overexpression or gain-of-function (GOF) missense mutations is implicated in the aetiology of multiple cancers^14–16. In multiple myeloma, the prevalent t(4;14) translocation causes aberrant NSD2 overexpression and increases H3K36me2 levels, which are thought to drive t(4;14)⁺ myelomagenesis^2,17–19. Two GOF mutations (E1099K and T1150A) that increase NSD2 catalytic activity are recurrently identified in haematological malignancies and are thought to promote oncogenesis^9,14–16,20. These GOF mutations are also sporadically detected in solid tumours^4,9,20,21, and NSD2 overexpression is common in solid tumours^3,14–16,22 (Extended Data Fig. 1b). For example, NSD2 promotes tumorigenesis in preclinical prostate cancer models^23–25, a result consistent with associations among H3K36me2, epithelial–mesenchymal transition (EMT) and cancer metastasis^26,27. NSD2 is also linked to pathogenic KRAS signalling in multiple myeloma and lung cancer^3–5. Thus, NSD2 is a compelling epigenetic target for drug development to treat diverse neoplasms²⁸.

Oncogenic KRAS mutations, which lock the expressed protein in the active state, are present in around 20% of all cancers and affect an estimated 3.4 million individuals every year^29,30. Pancreatic cancer and lung cancer—two challenging malignancies with poor responses to most systemic therapies—frequently have KRAS mutations³⁰. Notably, the selective KRAS^G12C inhibitors sotorasib (also known as AMG510) and adagrasib (also known as MRTX849) recently received approval by the US Food and Drug Administration for cancer treatment^31,32, and several other RAS inhibitors are in development^33–35. Although clinical outcomes for patients with KRAS^G12C lung adenocarcinoma (LUAD) treated with KRAS^G12C inhibitors are encouraging, drug resistance represents a significant challenge^30,36–38. As such, there is a need to identify clinically actionable targets that, alone or in combination with KRAS inhibitors, treat KRAS-driven tumours. Here we characterize highly potent and selective small-molecule NSD2 catalytic inhibitors and demonstrate their efficacy in preclinical models of KRAS-driven pancreatic cancer and lung cancer.

NSD2 accelerates KRAS^G12C-driven cancers

We previously found that transgenic overexpression of NSD2^E1099K rapidly accelerates tumorigenesis and malignant progression in vivo and decreases survival in a KRAS^G12D-driven LUAD mouse model⁴. As G12C is the most common KRAS oncogenic mutation in LUAD^29,30, we tested the impact of NSD2 hyperactivity in a KRAS^G12C-driven LUAD mouse model (named K^cP)³⁹ (Extended Data Fig. 1c,d). In this model, Kras^G12C expression and Trp53 homozygous deletion are induced through the intratracheal lavage of adenovirus expressing Cre recombinase (Ad-Cre)³⁹, with mutant mice succumbing to LUAD about 160 days after viral infection. NSD2 and H3K36me2 levels were higher in tumours than in normal lung cells (Extended Data Fig. 1e). Transgenic expression of NSD2^E1099K in K^cP mice further increased H3K36me2 signalling with concomitant H3K27me3 depletion (Extended Data Fig. 1f), a result consistent with the known antagonism of EZH2 by H3K36 methylation^7–13 (Extended Data Fig. 1a). Functionally, NSD2^E1099K expression substantially accelerated malignant progression, as demonstrated by an increased tumour burden and nodule number and a 60% reduction in overall median survival time in these mice (Extended Data Fig. 1g–i). Thus, like KRAS^G12D, NSD2 hyperactivity accelerates KRAS^G12C-driven malignancy in vivo⁴.

We also postulated that NSD2 promotes pancreatic ductal adenocarcinoma (PDAC) pathogenesis, as KRAS mutations are nearly universal³⁰ and NSD2 levels inversely correlate with patient survival in this malignancy (Fig. 1a). We therefore evaluated the impact of NSD2 hyperactivity on tumorigenesis in a KRAS^G12C-driven PDAC mouse model, designed in a similar manner as the canonical KPC mouse strain⁴⁰ (Fig. 1b). A conditional Kras^G12C allele that maintains physiological Kras expression from the endogenous locus until pancreas-specific activation was crossed with a Trp53^loxP/loxP strain to generate Kras^G12C/+;Trp53^loxP/loxP; Ptf1a^cre/+ mice (heterozygous K^cPC mice). These mice develop PDAC with 100% penetrance and have tumour-specific increased NSD2 expression (Fig. 1b and Extended Data Fig. 1j,k). In this aggressive PDAC model (K^cPC;Nsd2^E1099K), transgenic NSD2^E1099K expression further increased H3K36me2 levels that were accompanied by H3K27me3 depletion (Fig. 1c). Moreover, compared with K^cPC mice, K^cPC;Nsd2^E1099K mice rapidly progressed to late-stage PDAC (Fig. 1d), with phenotypes that included hyperplasia and reduced apoptosis and a substantial decrease in median postnatal survival from 91 to 35 days after birth (Fig. 1e and Extended Data Fig. 1l,m). These data suggest that catalytic inhibition of NSD2 may be therapeutically beneficial in PDAC and LUAD.

Fig. 1 |. NSD2 accelerates malignant PDAC and functional characterization of NSD2i.

a, NSD2 expression negatively correlates with survival in patients with PDAC. NSD2 immunohistochemistry chromogen staining intensity of patient samples (n = 74) was quantified and divided by high or low expression with a median cut-off value. b, Schematic of the generation of the K^CPC mouse model of PDAC expressing the catalytically hyperactive Nsd2^E1099K allele (K^CPC;Nsd2^E1099K). Pancreas-specific Cre-recombination of mutant alleles was mediated using the Ptf1a^cre strain. c, Western blot analysis with the indicated antibodies of representative lysates of PDAC tumour biopsy samples from K^CPC and K^CPC; Nsd2^E1099K mice at 4 weeks of age. Vinculin and total H3 were the loading controls. Arrowhead indicates NSD2. d, Representative macropathology and haematoxylin and eosin (H&E) staining of pancreatic tumours from K^CPC and K^CPC;Nsd2^E1099K mice at 4 weeks of age (representative of n = 5 per group). Scale bars, 5 mm (macro), 3 mm (H&E) or 50 μm (H&E zoom). e, Kaplan–Meier survival curves of K^CPC control mice (n = 12, median survival 91 days) and K^CP;Nsd2^E1099K mice (n = 5, median survival 35 days). f, Chemical structures of IACS-17596 and IACS-17817 (collectively termed NSD2i). g, Potency of the NSD2i compounds against a KMT panel. IC₅₀ values of in vitro methylation inhibition by the indicated compound for the indicated KMTs are based on n = 9 (NSD1–NSD3) or n = 6 (other KMTs) independent experiments. h, NSD2i treatment depletes H3K36me2 in a panel of PDAC and LUAD cell lines and human MSCs (hMSC). Western blot analysis with the indicated antibodies of whole cell extracts from indicated cell lines treated for 48 h with or without 100 nM IACS-17596. H3 and tubulin were loading controls. i, Cell viability dose–response curves (left) and IC₅₀ values (right) for the cell panel as in d treated with IACS-17596 at the indicated concentrations. Cell viability was determined after 7 days of treatment from three independent experiments. P values were determined using log-rank test (a,e). Data are the mean ± s.e.m. of the indicated number of biological replicates.

Clinical-grade NSD2 inhibitory compounds

There are currently no peer-reviewed descriptions of small-molecule NSD2 catalytic inhibitors that are potent, selective and active in vivo²⁸. However, a putative enzymatic inhibitor (KTX-1001) is currently being evaluated in a phase I clinical trial for multiple myeloma (ClinicalTrials. gov identifier: NCT05651932). We therefore synthesized a series of compounds from the patent⁴¹ that describes KTX-1001 and screened them for NSD2 methylation inhibitory potency (Supplementary Note 1). We identified one compound, IACS-17596, that inhibited NSD2 methylation of nucleosome substrates in vitro with a half-maximal inhibitory concentration (IC₅₀) of about 8.8 nM (Fig. 1f,g and Extended Data Fig. 2a). Medicinal chemistry efforts produced a second compound distinct from KTX-1001 (IACS-17817)⁴² (Supplementary Note 1) that also effectively inhibited NSD2 activity in vitro (IC₅₀ of about 19 nM; Fig. 1f,g and Extended Data Fig. 2a). In addition to inhibiting the NSD2 GOF E1099K and T1150A variants (Extended Data Fig. 2b), IACS-17596 and IACS-17817 (hereafter collectively termed NSD2i unless specified otherwise) were highly selective for NSD2 over related lysine methyltransferases (KMTs), including NSD1 and NSD3 (Fig. 1g and Extended Data Fig. 2a). Furthermore, both compounds did not show activity against a panel of broad kinase selectivity (Extended Data Fig. 2c and Supplementary Table 1).

Low-dose NSD2i treatment of MiaPaCa2 cells (a human PDAC cell line) depleted H3K36me2 levels within 24 h (Extended Data Fig. 2d–f). Such depletion was also observed in a panel of KRAS mutant cell lines (Fig. 1h) and mouse cancer cell lines (Extended Data Fig. 2g). The H3K36me2 loss is probably due to specific NSD2 targeting rather than other KMTs, as H3K36me2 levels did not further decrease in NSD2i-treated NSD2 knockout cell lines (Extended Data Fig. 2h). Functionally, NSD2i treatment reduced the viability of all cell lines in the panel, with IC₅₀ values ranging from 20 to 300 nM (Fig. 1i). NSD2i treatment also inhibited colony formation in PDAC cells (Extended Data Fig. 2i). In contrast to the mutant KRAS⁺ cancer cell panel, non-cancerous human mesenchymal stem cells (MSCs) were ≥25-fold less sensitive to NSD2i treatment (Fig. 1h,i). NSD2 is frequently overexpressed in cancer^4,14,22 (Extended Data Fig. 1b); accordingly, NSD2i treatment depleted H3K36me2 levels and reduced the viability of wild-type KRAS cancer cells driven by other oncogenic mutations (Extended Data Fig. 2j,k). We note that although NSD2i-mediated H3K36me2 depletion occurred within 24 h (Extended Data Fig. 2f), compromised cellular viability required sustained exposure of several days, a timing consistent with an epigenetic mechanism of action²⁸. Collectively, these data demonstrate that in diverse cancer cell lines, but not human MSCs, NSD2 generates the bulk of H3K36me2 and regulates cancer cell phenotypes through this histone modification.

Structural basis of NSD2i activity

To gain molecular-level insight into the mechanism of action of the NSD2i compounds, a high-resolution multidimensional solution nuclear magnetic resonance (NMR) spectroscopy strategy was used. First, 1D proton NMR spectrum competition experiments were performed with the unlabelled catalytic domain of NSD2 (NSD2(SET)) incubated with equimolar amounts of S-adenosylmethionine (SAM) and/or IACS-17596. Competitive binding of the two ligands was observed (Extended Data Fig. 3a), which indicated that IACS-17596 has a SAM-competitive inhibitory mechanism. This result is consistent with the fact that the NSD2i compounds contain an adenine purine analogue moiety (Fig. 1f). To map dynamic structural changes on NSD2(SET) after NSD2i engagement, backbone and selected methyl side-chain amino-acid resonance assignments for apo-NSD2(SET) and the IACS-17596–NSD2(SET) complex were obtained (Extended Data Fig. 3b). These analyses revealed significant chemical shift perturbations (CSPs) and signal broadening on the 3D structure of NSD2(SET) after the addition of IACS-17596 (Fig. 2a and Extended Data Fig. 3b). The main spectral changes clustered in three NSD2 functional regions: (1) the SAM-binding pocket, consistent with the competition experiments; (2) the catalytic channel; and (3) the autoinhibitory loop, which was highlighted by broadening peaks in this region (Fig. 2a,b and Extended Data Fig. 3a,b). Comparisons of the NSD2i–NSD2(SET) complex in solution with existing high-resolution NSD2 structures^43,44 suggested that the central pyridine and adenine analogue moieties of IACS-17596 mimic the ribose and adenine motifs of SAM to engage key NSD2 SAM-binding residues (Fig. 2b,c). Restricted stereochemistry at the 3-amino piperidine-pyridine groups of the NSD2i compounds further directed interactions with SAM-binding residues (Fig. 2a–c). Together, these findings indicate that SAM-mimetic chemistry underlies the ability of NSD2i compounds to directly compete with SAM to engage NSD2.

Fig. 2 |. Molecular characterization of the NSD2i compounds.

a, IACS-17596 localizes in both the SAM-binding pocket and the catalytic channel of NSD2. Amide CSPs (CSP^N/H) between NSD2 in its apo form and bound to IACS-17596 are mapped onto the surface of the stochiometric 1:1 complex in solution (Protein Data Bank (PDB) identifier: 9CVD). The coloured scale represents the size of significant CSP^N/H observed after IACS-17596 binding to NSD2(SET). Grey, prolines and other residues not detected in ligand-free and ligand-bound forms. Cyan, residues with backbone amides broadened beyond detection in the IACS-17596-bound state. IACS-17596, blue sticks model. b, IACS-17596 interacts with key NSD2 SAM-binding residues. Top, NSD2(SET)–IACS-17596 complex as in a. Bottom, an NSD2(SET)–SAM crystal structure (PDB: 5LSU) with residues mapped within 5 Å distance between non-hydrogen atoms of SAM and NSD2 indicated in blue. c, Comparative chemical structures of IACS-17596 (NSD2i) and SAM. The dotted box indicates NSD2i moieties that interact with the SAM-binding pocket and chemically mimic the SAM structure. d, Insertion of the fluorinated benzene ring of the NSD2i compound induces rearrangements in the enzyme catalytic channel. Bending of the catalytic tyrosine (Y1179) promotes inactive autoinhibitory loop conformations and H3 backbone docking site occlusion. The electron microscopy NSD2–SAM–rNUC (recombinant nuclesome) structure is adapted from PDB: 7E8D. e, Significant backbone CSPs observed in residues L1151 (red) and V1166 (yellow) despite their distance from the NSD2i-binding site. f, In vitro methylation assays (as in Fig. 1g) of wild-type (WT) and the L1151Q and V1166L (LQ/VL) mutant NSD2(SET) treated with 3 μM NSD2i compound or control (DMSO). Results are normalized to WT NSD2(SET) control levels. g, In vitro methylation assays as in f evaluating WT and the Q2030L and L2045V (QL/LV) mutant NSD1(SET) mutant. Data are the mean ± s.e.m. from three independent experiments (f,g). P values were determined using two-way analysis of variance (ANOVA) with Tukey’s testing for multiple comparisons.

In addition to a SAM-competitive mechanism, anchoring of IACS-17596 in the NSD2 SAM-binding pocket positioned the fluorinated benzene ring to insert into and induce rearrangements in the catalytic channel, which displaced the essential catalytic tyrosine residue (Y1179) (Fig. 2d). In the active conformation, the planar positioning of Y1179 formed the catalytic tunnel floor, where the H3K36 substrate side chain is positioned for methyl transfer (Fig. 2d). Introduction of a difluorinated benzene into the tunnel induced bent conformations for Y1179 that are predicted to obstruct H3K36 insertion into the catalytic centre (Fig. 2d). Binding of the compound further disrupted the catalytic core, which predisposed inactive autoinhibitory loop^43,44 conformations and induced occlusion of the H3 backbone docking site (Fig. 2d). We note that IACS-17596 and IACS-17817 interacted with similar key NSD2 residues, which suggests that they have similar inhibitory modes of action (Extended Data Fig. 3c,d). Thus, the high potency and selectivity of NSD2i compounds is probably due to multiple mechanisms that include SAM-competitive binding, steric-induced rearrangement of the catalytic tunnel and disruption of autoinhibitory loop dynamics.

The primary sequence that encompasses the SET domain of NSD2 is highly conserved, and NSD1 and NSD3 have equivalent regions (Extended Data Fig. 3e); however, the NSD2i compounds are highly selective for NSD2. Most of the NSD2 residues affected by ligand engagement cannot be mutated and analysed for their role in conferring selectivity because they are essential for catalytic function (Fig. 2a,b). However, L1151 and V1166 of NSD2, which show strong backbone CSPs, are not present in NSD1 or NSD3 and are far from key catalytic sites (Fig. 2e and Extended Data Fig. 3e). Notably, methylation activity remained intact in an NSD2(SET) mutant with the corresponding NSD1 and NSD3 residues (L1151Q and V1166L), but it was threefold less sensitive to the NSD2i compounds (IC₉₀ of 3 μM) (Fig. 2f). A reverse substitution in NSD1 (Q2030L and L2045V) rendered the NSD1(SET) mutant 25% more sensitive to the NSD2i compounds (Fig. 2g). Collectively, we speculate that the targeted selectivity of the NSD2i compounds is due to their direct engagement with specific primary sequences in the NSD2 catalytic core combined with allosterically induced tertiary dynamics that extend across the SET domain.

NSD2i-mediated epigenomic reprogramming

We and others have explored how NSD2 depletion affects the epigenome in development and oncogenesis^2,4,5,43,45. Moreover, the use of PROTAC-based tool compounds that target NSD2 has provided insights into the functions of NSD2 (refs. 25,46–50). However, investigating the consequences of specific and temporal loss of NSD2-catalysed H3K36me2 deposition on the cancer epigenome has previously not been possible. Here we treated MiaPaCa2 cells with or without an NSD2i compound (100 nM) and then performed integrated proteomic, epigenomic and transcriptomic analyses 1, 5 and 9 days after treatment (Fig. 3a). NSD2i treatment depleted H3K36me2 levels in MiaPaCa2 cells at all time points without affecting NSD2 abundance (Fig. 3b and Extended Data Fig. 4a). Quantitative mass spectrometry (qMS) profiling of bulk chromatin was performed to determine histone modification stoichiometry and dynamics in response to NSD2 inhibition. H3K36me2 levels were high in MiaPaCa2 cells and were found on >30% of all H3.1 and H3.3 nucleosomes (Fig 3c and Supplementary Table 2). NSD2i treatment reduced H3K36me2 levels by about 50% at day 1 compared with cells treated with dimethyl sulfoxide (DMSO) as a control. H3K36me2 levels were reduced by >80% at later time points (Fig. 3c and Supplementary Table 2). As expected, given the relationship between H3K27me3 and H3K36 methylation^7–13, NSD2i treatment resulted in an initial modest increase in H3K27me3 levels at day 1 and a full doubling in total H3K27me3 levels by days 5 and 9 (Fig. 3c and Supplementary Table 2). Other epigenetic changes included modest depletion of H3K36me1, more substantial depletion of H3K36me3 and H3K27me1, increased H3K27me2 and an unexpected general increase in acetylation levels (Fig. 3c and Supplementary Table 2). No other histone modifications that we analysed changed in a consistent or significant manner (Fig. 3c). Similar NSD2i-induced changes in the histone-modification landscape were observed in four additional KRAS-mutant cancer cell lines (Extended Data Fig. 4b and Supplementary Table 2). Collectively, these results show that in multiple cell types, NSD2 inhibition induces a substantial depletion in H3K36me2 levels and, after a lag, an increase in H3K27me3 levels that is accompanied by distinct additional epigenetic alterations.

Fig. 3 |. NSD2i treatment induces epigenetic and epigenome-wide reprogramming in PDAC cells.

a, Schematic of the integrated and temporal proteomic, epigenomic and transcriptomic experimental strategy to investigate NSD2i activity in MiaPaCa2 cells. D0, day 0; D1, day 1; D5, day 5; D9, day 9. b, Western blot analysis with the indicated antibodies of MiaPaCa2 whole cell extracts from the indicated conditions. H3 and tubulin were loading controls. See Extended Data Fig. 4a for additional histone modifications. c, qMS-based epigenetic profiling of the indicated histone modifications from cells treated as in a. Left, heatmap shows the log²[fold change] for each modification normalized to its relative abundance in controls (DMSO) at D1. Right, heatmap shows the stoichiometry indicated as the relative abundance for each histone modification relative to total H3.1 or H3.3 variants (Supplementary Table 2). qMS data are representatives from three independent biological replicates. d, Representative tracks of MS-normalized CUT&RUN signals from the qMS histone analysis, and read depth-normalized H3K4me3, NSD2 and NicE-seq signals on the indicated region of chromosome 2 for D1 control samples. Genes present in the analysed region are indicated at the bottom. e–g, NSD2i treatment causes H3K36me2 depletion and increased H3K27me3. e, MS-normalized H3K36me2 and H3K27me3 CUT&RUN signals in 10-kb genomic bins showing changes in the genomic distribution in the indicated conditions described in a. The colour scale differentiates genic versus intergenic regions of the genome and shows that changes occur at both. f, Representative tracks as in d of qMS-normalized H3K36me2 and H3K27me3 CUT&RUN signal as indicated from D5 samples comparing distribution changes in response to NSD2i treatment. g, MS-normalized CUT&RUN profiles of H3K36me2 and H3K27me3 over the averaged gene body (all genes) in the indicated conditions and time points show that NSD2i treatment causes H3K36me2 depletion with a concomitant increase in H3K27me3. TES, transcription end site; TSS, transcription start site.

To gain insight into how NSD2i treatment affects the epigenomic landscape, we performed automated cleavage under targets and release using nuclease (CUT&RUN)⁵¹ for NSD2, H3K36me2 and six additional canonical histone marks. In parallel, we also performed genome-wide chromatin accessibility profiling by nicking enzyme-assisted sequencing (NicE-seq)⁵² (Fig. 3d, Extended Data Fig. 4c and Supplementary Table 3). Compared with control conditions, H3K36me2 signals were progressively depleted over time across both genic and intergenic regions (Fig. 3e). The effect of NSD2i treatment on H3K27me3 distribution was bimodal, with the baseline intergenic signal retained and the emergence at days 5 and 9 of new H3K27me3 domains that were particularly intense in genic regions (Fig. 3e). Notably, baseline H3K27me3 peaks often intensified in signalling and spread to adjacent regions previously demarked by H3K36me2 (Fig. 3f and Extended Data Fig. 4d,e). Furthermore, across gene bodies, we observed an inverse relationship between H3K36me2 and H3K27me3 levels that intensified over time (Fig. 3g and Extended Data Fig. 5a). This relationship was also observed across intergenic regions (Extended Data Fig. 5b). The genomic distribution of NSD2, which under control conditions was highly correlated with H3K36me2 (R = 0.79, P < 2.2 × 10⁻¹⁶), was not significantly altered by NSD2i treatment (Extended Data Fig. 5c). This finding indicated that NSD2 remains bound to chromatin irrespective of whether its catalytic activity is blocked. The distribution of most of the other histone marks analysed were unaltered by NSD2 inhibition (Extended Data Fig. 5c). However, H3K36me3 was attenuated across gene bodies, as expected given its 60–70% depletion as measured by qMS (Extended Data Fig. 5c and Supplementary Table 2). Collectively, these results show that NSD2i exposure induces substantial reprogramming of distinct active and repressive marks across the genome, as highlighted by the invasion of H3K27me3 in intergenic and genic regions previously populated by H3K36me2.

NSD2i treatment of MiaPaCa2 cells did not initially result in transcriptomic changes. However, by days 5 and 9, there were significant differentially expressed genes (DEGs), with considerably more downregulated than upregulated DEGs (Fig. 4a,b and Supplementary Table 3). This result is consistent with the idea that loss of H3K36me2 impairs transcriptional activation¹⁵. Comparisons of the differential transcriptomes with the Hallmark gene set signature collection identified downregulation of KRAS signalling and EMT as two of the six functional pathways significantly altered (all suppressed) at the different time points (Fig. 4c and Supplementary Table 4). Moreover, gene set enrichment analysis (GSEA) of two PDAC-activating profiles^53,54 showed significant downregulation in both signatures at all time points in response to NSD2i treatment (Fig. 4c). Similar NSD2i-induced gene expression changes were observed in transcriptomic analyses of KP2 and H1373 cell lines (Extended Data Fig. 6a and Supplementary Table 4). Thus, NSD2i treatment of three distinct mutant KRAS⁺ cells leads to the repression of select gene expression programs that are predicted to contribute to oncogenesis.

Fig. 4 |. NSD2i treatment suppresses oncogenic programming through H3K27me3 upregulation at H3K27me3 legacy loci.

a,b, NSD2i treatment inhibits gene expression. a, Differentially downregulated DEGs (dDEGs) and upregulated DEGs (uDEGs) from RNA-seq analysis of MiaPaCa2 cells. The absolute fold change cut-off value was 1.5 and the adjusted P cut-off value was 0.05. b, D5 volcano plot of data in a. c, NSD2i treatment suppresses oncogenic gene expression programs. GSEA of MiaPaCa2 cells with or without NSD2i-induced differential transcriptomes at the indicated time points. Signatures with significant changes at all time points and PDAC-activated signatures are shown (Supplementary Table 4). NES, normalized enrichment score. d, GSEA of PRC2 target signature changes like c at the indicated time points. e–g, NSD2i treatment selectively suppresses the expression of indicated gene categories. Genes in functional and cognate control groups have similar expression levels (Extended Data Fig. 6b). PRC2 targets n = 469, PRC2 neighbours n = 738, PRC2 enhancer n = 1,000, pancreatic neoplasia n = 304, KRAS signalling n = 107, EMT n = 125 genes versus the matching number of controls genes. Violin plots show boxes with 25th to 75th percentiles, whiskers with minimum to maximum values, and the centre line the mean. e, NSD2i treatment inhibits PRC2 target gene expression significantly more than control genes at D5 and D9. f, NSD2i treatment inhibits the expression of genes neighbouring PRC2 targets (within 400 kb) significantly more than control genes as in e. g, Enhancer-dependent genes are differentially downregulated by NSD2i treatment relative to cognate control genes as in e. h, Violin plots of MS-normalized H3K27me3 CUT&RUN signals of genes in indicated groups, conditions and time points. i, Difference in mean MS-normalized H3K27me3 CUT&RUN signals from h between the indicated signature and cognate control signatures at the indicated conditions and time points. j, Fold change of mean MS-normalized H3K27me3 CUT&RUN signals at the indicated target over control signatures in vehicle-treated samples for each time point. The dotted line indicates the expected H3K27me3 signal, values above which indicate H3K27me3 enrichment for the target over the control signature. P values were determined using two-sided Wald test (b), two-sided permutation test (c,d), or two-sided Wilcoxon signed-rank paired test (e–h).

NSD2i treatment induced a significant decrease in expression of the top 500 PRC2 target genes in MiaPaCa2 cells (Fig. 4d and Supplementary Table 4), a result that potentially links the epigenomic and transcriptomic findings. To test whether this effect is specific, we assembled a control gene set signature with a matching number of equally expressed genes as the PRC2 target signature (Extended Data Fig. 6b). At both days 5 and 9, NSD2i treatment suppressed the expression of genes in the PRC2 target signature but not the control signature (Fig. 4e; insufficient DEGs precluded day 1 analysis). Considering the temporal and genomic spread of H3K27me3 marks in response to NSD2i treatment (Fig. 3e,f), we postulated that gene proximity to PRC2 targets may underlie an additional level of H3K36me2-dependent transcriptional regulation. Indeed, neighbouring genes (defined as within 400 kb) of PRC2 targets were repressed significantly more than genes from a cognate control signature (Fig. 4f and Extended Data Fig. 6b). Transcriptional enhancers are influenced by the H3K36me2–H3K27me3 balance⁵⁵. Accordingly, we observed that the expression of genes with high enhancer dependency were more significantly inhibited by NSD2i treatment than a cognate control gene set (Fig. 4g and Extended Data Fig. 6b). These data provide support for a link between H3K27me3 dynamics and the selective susceptibility of genes to NSD2 inhibition.

We next assessed whether there were differences in H3K27me3 levels at genes in NSD2i-regulated GSEA signatures compared with their cognate control gene sets (Extended Data Fig. 6b). NSD2i treatment induced significantly greater changes in H3K27me3 levels across all time points in the pancreatic neoplasia, KRAS signalling and EMT gene signatures than their respective control gene sets (Fig. 4h,i). Furthermore, regions that gained H3K27me3 marks after NSD2i treatment (for example, the EMT signature) had higher levels of H3K27me3 at baseline than control signatures (Fig. 4j). For example, under control conditions at day 1, the mean H3K27me3 signal at EMT signature genes was about threefold greater than that of the cognate control group, with similar trends observed for the other signatures (Fig. 4j). The pattern for H3K36me2 was generally the inverse of H3K27me3 (Extended Data Fig. 6c). Chromatin accessibility was also lower in the NSD2i-regulated signatures than control GSEA signatures, a result consistent with the observed histone methylation changes (Extended Data Fig. 6d). Collectively, these results highlight epigenomic features of genes at baseline that are selectively repressed by NSD2i treatment compared with non-regulated genes. These genes exhibit the following characteristics: (1) increased H3K27me3 levels; (2) reduced promoter chromatin accessibility; and/or (3) high enhancer dependency. We speculate that many genes repressed by NSD2i treatment may represent H3K27me3 legacy loci silenced in normal differentiated tissue. In the transformed cancer state, active transcription of these genes may rely on NSD2-catalysed H3K36me2 to limit H3K27me3 to maintain an open chromatin configuration (Discussion).

NSD2 inhibition as a cancer therapeutic

Inhibitors of oncogenic KRAS mutations have shown encouraging results in the clinic. Among this compound class, so far, only KRAS^G12C-selective inhibitors (for example, sotorasib (hereafter referred to as KRASi) and adagrasib) are approved for clinical use in oncology indications^31,56. However, owing to drug resistance, it is useful to identify clinically actionable targets that cooperate with KRAS inhibitors to improve treatment outcomes^30,31,38. Considering that the NSD2i compounds suppress KRAS-associated and other oncogenic transcriptional programs (although KRASi does not affect NSD2 or H3K36me2 levels (Extended Data Fig. 7a)), we postulated that combining an NSD2i compound with KRASi may be therapeutically beneficial. To test this idea, colony formation was assessed in patient-derived PDAC and LUAD cells treated with an NSD2i compound (0–100 nM) and KRASi (0–25 nM). The combination treatment led to Bliss–Loewe synergy scores of 32.5 and 21.6, respectively, whereby a score of >10 indicates synergy⁵⁷ (Fig. 5a and Extended Data Fig. 7b).

Fig. 5 |. Targeting NSD2 is an actionable therapeutic vulnerability in KRAS-driven cancers.

a, NSD2i and KRASi synergize in PDAC PDX cells. Bliss–Loewe score was calculated from colony formation assays. b, Scheme of the experimental design to treat subcutaneous PDX grafts. At the specified time point, animals were treated with or without NSD2i (100 mg kg⁻¹, daily), with or without KRASi (sotorasib, 10 mg kg⁻¹, daily) or with vehicle control. c,d, Tumour volume quantification for PDAC KRAS^G12C;Trp53^R282W (c) and LUAD KRAS^G12C;Trp53^R273C (d) PDX xenografts in NSG mice (n = 5 mice per group) as in b. Arrow indicates treatment commencement. 17596, IACS-17596; 17817, IACS-17817. e, Scheme of the experimental design to assess indicated treatments efficacies in the K^CPC PDAC model. At the specified time point, animals were treated with or without NSD2i (100 mg kg⁻¹, daily), with or without KRASi (sotorasib, 10 mg kg⁻¹, daily) or with vehicle control as indicated until the end point criteria were met. f, Western blots with the indicated antibodies and representative K^CPC PDAC biopsy lysates at 2 weeks after treatment. Vinculin and H3 were the loading controls. g, Kaplan–Meier survival curves of K^CPC control (n = 12) and indicated treatments (n = 7 per group) as in e. Gemcitabine (100 mg kg⁻¹, every three days) is shown for comparison. Median time of survival after treatment is indicated. h, Scheme of the experimental design to assess indicated treatment efficacies in the K^C/CP LUAD mouse model as in e. i, Western blots like f of representative K^C/CP lung biopsy lysates. j, Kaplan–Meier survival curves of K^C/CP mice treated with vehicle control or the indicated inhibitors (n = 7 per group). Median time of survival after treatment is indicated. k, Quantification of tumour nodules per lung at the humane end point as in j (n = 7 per group). l, Tumour fitness score derived from scRNA-seq analysis from indicated microdissected K^C/CP tumours (Extended Data Fig. 9c). Control n = 987, NSD2i n = 671, KRASi n = 2185, NSD2i + KRASi n = 94 cells. m, scRNA-seq data of GSEA profiling of the indicated signatures and conditions. Circle sizes, adjusted P; colour scale, NES value. P values (indicated on the charts) were determined using two-way ANOVA with Tukey’s testing for multiple comparisons (c,d,k), log-rank test (g,j) or two-sided permutation test (m). Data are the mean ± s.e.m. of the indicated number of biological replicates. Boxes show the 25th to 75th percentiles, whiskers the minimum to maximum values, and the centre line the median (k,l).

The NSD2i compounds showed acceptable drug-like absorption, distribution, metabolism and excretion (ADME) and pharmacokinetic properties. That is, 100 mg kg⁻¹ intraperitoneal daily dosing of each compound in mice led to on-target effects and was well tolerated (Extended Data Figs. 7c–g and 8a). For example, NSD2i treatment did not affect mouse body weight or compromise the integrity of several organs (Extended Data Fig. 7f,g). Moreover, mice treated with an NSD2i compound exhibited normal complete blood counts and blood chemistries, with no signs of leukopaenia, neutropaenia, anaemia, thrombocytopaenia or compromised liver or kidney function (Extended Data Fig. 8a). Thus, the NSD2i compounds seem to be nontoxic to normal cells and tissues and have a favourable therapeutic window.

We therefore tested the NSD2i compounds alone and combined with KRASi in vivo in a KRAS^G12C-positive PDAC model and a patient-derived xenograft (PDX) LUAD model (Fig. 5b). In each disease model, NSD2i treatment led to strong single-agent tumour growth inhibition (Fig. 5c,d). KRASi alone inhibited PDX growth, with faster onset kinetics than the NSD2i compounds. However, by 2–3 weeks, tumours were less responsive to KRASi, which may reflect the development of treatment resistance^30,31,38 (Fig. 5c,d). Notably, combining either NSD2i compound with KRASi led to rapid and sustained tumour regression, more so than with either agent alone (Fig. 5c,d), a result consistent with their synergistic effects observed in cell culture. Animal weight was maintained for the duration of the experiments (Extended Data Fig. 8b,c).

We also tested single-agent and combination treatment modalities in advanced-stage autochthonous K^cPC PDAC mouse models (Figs. 1 and 5e). PDAC tumour lysates from mice treated for 2 weeks showed on-target activity. The NSD2i compounds, irrespective of KRASi, depleted H3K36me2, whereas KRASi, irrespective of NSD2i, blocked ERK1 and ERK2 (ERK1/2) phosphorylation (Fig. 5f). Furthermore, animal weight was maintained throughout all treatments, which suggests that the NSD2i compounds have minimal off-target toxicity, both alone and in combination with KRASi (Extended Data Fig. 8d). In survival studies, NSD2i treatment extended median survival by about 150% compared with control conditions or standard chemotherapy (Fig. 5g). KRASi treatment was slightly more effective in extending survival than NSD2i treatment (Fig. 5g). Notably, the NSD2i–KRASi combination extended median survival by about 280% compared with control animals, a result that exceeded the 225% predicted for a purely additive mechanism (Fig. 5g). The efficacy of the NSD2i compounds was also investigated in a hyperaggressive K^c/cP LUAD mouse model, in which both alleles express KRAS^G12C. Therapeutic intervention was initiated in K^c/cP animals with late-stage disease to recapitulate how patients often present to the clinic (Fig. 5h). On-target activity of NSD2i and KRASi treatments was observed in microdissected lung tumour lysates (Fig. 5i), and weight was maintained during treatment (Extended Data Fig. 8e). Single-agent treatments with NSD2i or KRASi compounds increased median survival by about 200% compared with vehicle control. By contrast, combination therapy increased median survival by >450% (Fig. 5j), a result consistent with a therapeutically synergistic interaction between the NSD2i and KRASi compounds.

LUAD in humans typically presents with one or a few primary neoplastic masses. By contrast, KRAS-driven LUAD in mice results in the formation of many primary tumour nodules. Therefore, tumour number counts in these mice are a useful metric to assess disease progression and treatment efficacy. In our K^c/cP model, an average of about 100 tumour nodules per mouse were observed at the start of the treatment regimen, a number that remained largely unchanged by the humane end point in control mice (Fig. 5k). Treatment with NSD2i or KRASi as single agents eliminated approximately 50% and 75% of individual tumours, respectively (Fig. 5k). Combination therapy eradicated nearly all tumour masses, with growth of a single large nodule ultimately leading to lethality (Fig. 5k).

To gain insight into our therapeutic interventions, single-cell RNA sequencing (scRNA-seq) was performed on fresh, microdissected LUAD tumour biopsy samples from K^c/cP mice treated for 3 weeks with one of the four different treatment regimens (Extended Data Fig. 9a,b). Quantification of overall oncogenic fitness, using a signature that reflected LUAD tumour expansion and evolution⁵⁸, demonstrated a decreased score in response to NSD2i, KRASi and combination therapies (Fig. 5l and Extended Data Fig. 9c). GSEA evaluation of scRNA-seq data showed that NSD2i treatment suppressed the same functional signatures (for example, EMT) as those observed in cancer cells (Figs. 4c and 5m and Supplementary Table 4). We also observed further suppression of KRAS-dependent and EMT pathways with combined, compared with individual, treatments (Fig. 5m). Notably, NSD2i treatment induced significantly greater transcriptional changes in cancer compared than in noncancer epithelial cells, as assessed using the Bhattacharyya distance (Extended Data Fig. 9d). This differential effect on cancer cell transcription was also evident for genes from the key NSD2i-suppressed oncogenic signatures (Extended Data Fig. 9e). These data confirm the favourable safety profile of the NSD2i compounds. Finally, independent of KRASi exposure, NSD2i treatment induced tumour cells to express an epithelial alveolar gene signature (Fig. 5m). Collectively, these data suggest that in an in vivo setting, NSD2i-mediated epigenomic reprogramming facilitates the reversion of cancer cell transcriptomes from a malignant state towards a differentiated epithelial state and potentially contribute to durable antitumour responses.

Discussion

NSD2 has long been considered a promising epigenetic target for pharmacological intervention in oncology^16,28. An NSD2 catalytic activity inhibitor was recently reported to have high in vitro potency⁵⁹; however, whether this compound inhibits NSD2 or H3K36me2 generation in cells or in vivo or exhibits selectivity over related enzymes like NSD1 is unknown⁵⁹. However, before this study, to our knowledge, no bona fide clinical-grade small-molecule catalytic inhibitor of NSD2 had been reported. To address this gap, we synthesized IACS-17596 and IACS-17817—two compounds related to KTX-1001 (PubChem CID 168429488), a drug in early-stage clinical trial evaluation (NCT05651932). IACS-17596 and IACS-17817 displayed high in vitro and in vivo potency and selectivity towards NSD2. Dynamic structural analyses revealed that both compounds use a modified bi-substrate engagement mechanism that is proposed to be optimal for inhibiting methyltransferases⁶⁰. A similar design strategy may facilitate the development of selective inhibitors against other promising KMTs such as NSD3 (ref. 11).

NSD2 activity increases during malignant progression in many human cancers and multiple cancer mouse models^{4,22–24,61,62} (Extended Data Fig. 1). We postulate that the pervasive and recurrent increase in NSD2 activity across diverse malignancies is linked to its ability—through H3K36me2—to shift select genomic loci from silenced to permissive chromatin states. We suggest that the downstream activation of mutant KRAS transcriptional targets is constrained by residual chromatin barriers that stem from the legacy of H3K27me3-mediated facultative heterochromatin at these loci (Extended Data Fig. 10a). This finding is consistent with our observations that NSD2i-regulated genes have significantly higher baseline levels of H3K27me3 than control genes (Fig. 4h,j). During mutant KRAS-driven carcinogenesis, pathologic NSD2-mediated H3K36me2 deposition may reduce H3K27me3 levels at target loci, which renders the underlying chromatin more accessible to KRAS-dependent pro-oncogenic transcription factors (Extended Data Fig. 10a). This hypothesis posits that the transcriptional targets most sensitive to NSD2i-mediated repression are developmental genes that are normally silenced during differentiation but reactivated by aberrant NSD2–H3K36me2 activity during the de-differentiation associated with tumorigenesis (Extended Data Fig. 10a). Thus, NSD2i-mediated H3K36me2 depletion at oncogenic gene targets may promote facultative heterochromatin reformation, reduce transcriptional output and ultimately suppress tumour growth (Extended Data Fig. 10a). It is unclear why some genes are more susceptible to NSD2i treatment than others. We speculate that NSD2i-repressed loci, which have inherently higher baseline H3K27me3 levels (Fig. 4i), represent H3K27me3 legacy loci. That is, these regions retain an epigenetic memory of their ‘differentiated’ state, which manifests as a modest, heterochromatin-like barrier to transcriptional activation by oncogenic signalling. The activation of such genes may become dependent on NSD2-catalysed H3K36me2, which limits H3K27me3 deposition. This in turn renders them particularly sensitive to reverting to their native, low-accessibility chromatin state after NSD2 inhibition (Extended Data Fig. 10a). We propose that other cancer types with increased NSD2 activity may be susceptible to NSD2i treatment through chromatin-based mechanisms similar to those described here (Extended Data Fig. 10b). Collectively, our results shed light on the mechanistic basis of NSD2 inhibition as a clinically actionable epigenetic dependency of KRAS-driven cancers, with potential application in a wide range of oncology indications.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-025-09299-y.

Methods

Cell lines and cell assay

MiaPaCa2 (American Type Culture Collection (ATCC), CRL-1420, male, 65 years old, PDAC), Panc1 (ATCC, CRL-1469, male, 56 years old, PDAC), T3M4 (RIKEN, RCB1021, male, age not reported, PDAC), PaTu8988S (DSMZ, ACC 204, female, 64 years old, PDAC), 293T (ATCC, CRL-3216, female, embryonic kidney), 4T1 (ATCC, CRL-2539, mouse mammary cancer), B16F10 (ATCC, CRL-6475, mouse melanoma) KP LUAD (gift from M. Winslow) and KPC PDAC (gift from M. Winslow) cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM, Gibco) supplemented with 10% FBS (Gibco), 1% sodium pyruvate (Gibco) and 1% penicillin–streptomycin (Gibco). NCI-H1650 (ATCC, CRL-5883, male, 27 years old, NSCLC), NCI-H1975 (ATCC, CRL-5908, female, age not reported, NSCLC), HCC827 (ATCC, CRL-2868, female, 39 years old, NSCLC), PC9 (RIKEN, RCB4445, sex and age not reported, NSCLC), NCI-H2228 (ATCC, CRL-5935, female, age not reported, NSCLC), NCI-H3122 (CYTION, 300484, male, age not reported, NSCLC), NCI-H520 (ATCC, HTB-182, male, age not reported, NSCLC), NCI-H1648 (ATCC, CRL-5882, male, 39 years old, NSCLC), NCI-H2110 (ATCC, CRL-5924, sex and age not reported, NSCLC), KP2 (HSRRB, JCRB0181, female, 65 years old, PDAC), Panc05.04 (ATCC, CRL-2557, female, 77 years old, pancreatic cancer), PSN1 (ATCC, CRL-3211, male, 36 years old, PDAC), NCI-H1373 (ATCC, CRL-5866, male, 56 years old, NSCLC), A549 (ATCC, CCL-185, male, 58 years old, NSCLC) and Corl23 (ECACC, 92031919, male, 62 years old, NSCLC) cells were cultured in RPMI 1640 medium (Gibco) supplemented with 10% FBS (Gibco), 2 mM l-glutamine (Gibco) and 1% penicillin–streptomycin (Gibco). G361 (ATCC, CRL1424, male, 31 years old, melanoma) cells were cultured in McCoy’s 5a medium modified (Gibco) supplemented with 10% FBS (Gibco) and 1% penicillin–streptomycin (Gibco). Human MSCs (gift from N. Riggi)⁶³ were cultured in MSC basal medium (ATCC, PCS-500–030) supplemented with reagents from a MSC growth kit (ATCC, PCS-500–041) and 10% FBS (Gibco). All cell lines were tested negative for mycoplasma using a Lonza mycoplasma detection kit (Lonza, LT07418). All cells were cultured at 37 °C in a humidified incubator with 5% CO₂. For the cell viability assays, 500 cells were seeded in 96-well plates with 3 biological replicates. Cells were treated with IACS-17596 at the indicated concentrations or DMSO. The cell culture medium was changed every 3 days. Viability was measured using PrestoBlue (ThermoFisher). The reactions were detected by luminescence and the relative activity of the enzyme was calculated. IC₅₀ values were determined using a four-parameter log dose, nonlinear regression analysis, with sigmoidal dose–response (variable slope) curve fit using Prism (v.9.5) software. For the colony formation assays, 200 cells were seeded in 6-well plates with 3 replicates. Cells were treated with NSD2i (1 μM) or DMSO. The cell culture medium was changed every 3 days. After 2 weeks of treatment, cells were fixed with 100% ethanol and stained with 0.5% crystal violet in methanol and colonies were counted.

Generation of clonal NSD2 knockout mouse cell lines

Transfections were performed with polyethylenimine in 293T cells. Stable cell lines were generated using lentiviral transductions. 293T cells were co-transfected with lentiviral plasmid, pCMV-Δ8.2 (Addgene, 8455) and pCMV-VSVg (Addgene, 8454) in a ratio of 5:4:1 by mass. At 48 h after transfection, target cells were transduced with 0.45-μm filtered viral supernatant supplemented with 4 μg ml⁻¹ polybrene. Cells were selected 48 h after medium replacement with 2 μg ml⁻¹ puromycin (ThermoFisher A1113802). To deplete NSD2 in cells, a CRISPR–Cas9 system was used with lentiCRISPRv2 puro (a gift from B. Stringer, Addgene, 98290) with the following sgRNAs: sgControl, 5′-CTTCGAAATGTCCGTTCGGT-3′; sgmNSD2, 5′-TCAGGGTCTCACAATTGGGC-3′ (obtained from ref. 64). Next, cells were detached from plates using trypsin and passed through a 35-μm cell strainer. Single-cell suspensions were stained with FxCycleTM Violet stain (1:1,000, ThermoFisher) according to the manufacturer’s instructions. All samples were gated on viable cells followed by exclusion of cell doublets using BD FACS Diva software (BD Biosciences). Cells were sorted into 96-well plates (Falcon) containing 100 μl fully supplemented DMEM using a BD Biosciences Aria cell-sorting platform with a 100 μm nozzle.

Expression and purification of recombinant proteins

With the exception of the MLL1 complex (active motif) and the PRC2 complex (active motif), the NSD1(SET) and mutants (residues 1853–2093, National Center for Biotechnology Information (NCBI) sequence: NC_000005.10), NSD2(SET) and mutants (residues 958–1365, NCBI sequence: NC_000004.12), NSD3(SET) (residues 1021–1320, NCBI sequence: NC_000008.11), ASH1L(SET) (residues 1980–2564, NCBI sequence: NC_000001.11), SETD2(SET) (residues 1418–1714, NCBI sequence: NC_000003.12), SETDB1(SET) (residues 590–1291, NCBI sequence: NC_000001.11), SUV39H1(SET) (residues 82–412, NCBI sequence: NC_000023.11), SUV39H2(SET) (residues 61–410, NCBI sequence: NC_000010.11), SUV420H1(SET) (residues 64–312, NCBI sequence: NC_000011.10) and SUV420H2(SET) (residues 2–248, NCBI sequence: NC_000019.10) were cloned into pGEX-6P-1 separately, and Escherichia coli BL21 cells were transformed with the respective expression vectors and cultivated in LB medium (10 g l⁻¹ tryptone, 5 g l⁻¹ yeast extract and 10 g l⁻¹ NaCl) supplemented with 0.1 mM isopropyl 1-thio-β-d-galactopyranoside (Sigma) at 18 °C for 16–20 h. Cells were lysed using a sonicator, lysates were cleared by centrifugation at 14,000 r.p.m. for 1 h and the supernatants were incubated with glutathione sepharose (GE Healthcare) for purification. Recombinant proteins were eluted in 10 mM reduced glutathione (Sigma). Protein concentrations were measured using a Pierce Coomassie Plus assay.

In vitro methylation reactions

Methylation reactions on nucleosomes were performed as previously described⁴³. In brief, 100 nM recombinant enzymes mixed with 250 nM mononucleosome (EpiCypher) in reaction buffer (50 mM Tris pH 8.0, 20 mM KCl, 5 mM MgCl₂ and 10% glycerol) were treated with 20 μM SAM and NSD2i, and the mixture was incubated at 30 °C for 3 h. Next, a MTase-Glo methyltransferase assay kit (Promega) was used to detect the reaction intensity and quantified.

In vitro determination of IC₅₀ for inhibitors

A MTase-Glo methyltransferase assay kit (Promega) was used to measure the IC₅₀ values for the NSD2i compounds for a panel of KMTs. Specifically, we established an assay in 10 μl reaction mix containing 25 nM KMT enzyme (or enzyme complex), 0.5 μM SAM, 100 nM mononucleosomes (EpiCypher), MTase-Glo reagent (1×) and 3-fold serial dilutions of inhibitor from 0 to 30 μM in reaction buffer (50 mM Tris pH 8.0, 20 mM KCl, 5 mM MgCl₂ and 10% glycerol) arrayed in a white 384-well microplate (Corning). Each independent biochemical reaction was performed 9 times (NSD1, NSD2 and NSD3) or 6 times (ASH1L, SETD2, MLL1, SUV39H1, SUV39H2, SETDB1, PRC2, SUV420H1 and SUV420H2) and incubated for 3 h at 30 °C. Subsequently, 10 μl MTase-Glo detection solution was added and incubated for 1 h at room temperature. The luminescence of the reactions was measured and the relative activity of enzyme was calculated. IC₅₀ values were determined using a four-parameter log dose, nonlinear regression analysis, with sigmoidal dose–response (variable slope) curve fit using Prism (v.9.5) software.

Western blot analysis

For western blot analysis, 1 million cells were lysed in 200 μl SDS buffer and boiled at 100 °C for 20 min. Protein samples were resolved by SDS–PAGE and transferred to a PVDF membrane. The following antibodies were used (at the indicated dilutions): H3K36me2 (ThermoFisher, 701767; 1:1,000); H3K36me3 (ThermoFisher, MA5–24687; 1:1,000); H3K27me3 (ThermoFisher, MA5–33081; 1:1,000); H3 (EpiCypher, 13–0001; 1:20,000); β-tubulin (CST, 2128; 1:1,000); KRAS (Abcam, ab275885; 1:1,000); vinculin (CST, 13901; 1:1,000); pERK1/2 (CST, 4370; 1:1,000); ERK1/2 (CST, 4695, CST 9102; 1:1,000); NSD2 (CST, 65127; 1:1,000)²; H4K20me (CST, 5737; 1:1,000); H3K9me3 (Abcam, ab8898; 1:1,000); H3K4me3 (EpiCypher, 13–0028; 1:1,000); H3K27ac (Abcam, ab4729; 1:1,000); H3K9ac (CST, 9649T; 1:1,000); H3K14ac (Abcam, ab52946; 1:1,000); H3K23ac (Abcam, ab4729; 1:1,000); secondary anti-mouse (CST, 7076; 1:10,000); and anti-rabbit (CST, 7074; 1:10,000) peroxidase-conjugated antibodies.

Protein expression for structural studies

The NSD2(SET) (residues 984–1206) sequence was incorporated into a pGEX-6P expression vector and transformed into the BL21 DE3 E. coli strain. Expression was carried out at 20 °C for 16–20 h in LB (competition experiments) or M9 medium (backbone and side-chain assignment and protein–ligand structure elucidation) containing 1 g l⁻¹ of ¹⁵NH₄Cl as the sole source of nitrogen and 2.5 g l⁻¹ of U-¹³C glucose as the source of carbon for the uniformly double ¹³C,¹⁵N labelled NSD2(SET) variant. For ILV-CH₃ selectively methyl labelled and perdeuterated ²H,¹³C and ¹⁵N samples, M9 medium was prepared in 99.9% D₂O with 2.5 g l⁻¹ of U-²H,¹³C glucose and 1 g l⁻¹ of ¹⁵NH₄Cl with the addition of ILV labelled precursors (110 mg α-ketoisovaleric acid and 65 mg α-ketobutyric acid) 30 min before induction and for 4 h of expression time. Following expression, the cells were collected, resuspended in 50 mM Tris pH 7.5, 300 mM NaCl and 1 mM TCEP and lysed using a French press. The supernatant was injected into a GSTrap HP 5 ml column and eluted with buffer containing 50 mM Tris pH 7.5, 300 mM NaCl, 1 mM TCEP and 10 mM reduced glutathione. The fractions containing NSD2(SET) were collected and combined with GST-PreScission. The mixture was then dialysed overnight at 4 °C against storage buffer composed of 50 mM Tris pH 7.5, 150 mM NaCl and 1 mM TCEP. The final purification was achieved using a GSTrap HP 5 ml column, and the flow-through was collected and concentrated with an Amicon 3000 MWCO to the desired concentration. Ligand stocks of 150 mM were freshly prepared in DMSO-d₆ before use.

Solution NMR assignment

All NMR data were acquired using 800 and 950 MHz Bruker Avance NEO MHz spectrometers equipped with 5 mm TCI cryogenic probes. Measurements were performed on samples with NSD2(SET) concentrations of 100–250 μM containing 1% D₂O (v/v), with the final concentration of DMSO-d₆ never exceeding 0.8% (v/v). The competition experiment for elucidating the ligand-binding mechanism was conducted using a set of 1D ¹H NMR experiments for the titrations series, with 1:1:1 molar ratio of NSD2(SET)–SAM–IACS-17596 and a 1:1 titration of NSD2(SET)–IACS-17596. The CSPs were determined with 2D ¹H-¹⁵N HSQC-TROSY experiments between the ligand-free and IACS-17596-bound or P1-bound forms and calculated using the formula ${\text{CSP}}_{\text{N}/\text{H}}={[\delta {\text{H}}^2+{(0.14\times \delta \text{N})}^2]}^{1/2}$ (ref. 65). Sequence-specific backbone resonance assignments of ¹H^N, ¹⁵N, ¹³C, ¹³Cα and ¹³Cβ resonances for NSD2(SET) in the apo and IACS-17596-bound forms were achieved on a 100–200 μM perdeuterated U-²H,¹³C,¹⁵N-ILV(–¹³CH₃) sample using TROSY-based triple resonance 3D experiments on HNCA, HNcoCA, HNCO, HNcaCO, HNCACB, HNCB and HNcoCACB, with further support from H^N-H^N connectivities found in 3D ¹⁵N-edited NOESY-HMQC (mixing time of 100 ms for ¹³C and ¹⁵N and 200 ms for perdeuterated U-²H,¹³C and ¹⁵N labelled samples for both NSD2(SET) ligand-free and the NSD2(SET)–IACS-17596 complex) spectra. The backbone (ψ and φ) and side chain (χ¹) torsion angles were determined using the assigned chemical shifts and the Talos-N algorithm⁶⁶. The side-chain assignments of methyl groups for ILVAMT amino acids for NSD2(SET) ligand-free and the NSD2(SET)–IACS-17596 complex were accomplished through high-resolution through-bond correlation spectra 3D (H)CCH-TOCSY (mixing times of 5.6 ms and 16.4 ms on U-¹³C and ¹⁵N labelled samples), 3D ¹⁵N-edited NOESY-HMQC (mixing time of 100 ms on U-¹³C,¹⁵N and 200 ms on perdeuterated U-²H,¹³C,¹⁵N-ILV labelled samples) and 3D ¹³C-edited NOESY-HSQC (mixing time of 100 and 200 ms on U-¹³C,¹⁵N) centred on the aliphatic region. The H^N-H^N and H^N-CH₃ (ILV) connectivities were analysed from 3D ¹⁵N-edited NOESY-HMQC (mixing time of 200 ms) and 3D ¹³C-edited NOESY-HSQC (mixing time of 100 ms on ILV perdeuterated U-²H,¹³C,¹⁵N NSD2(SET) samples in ligand-free and ligand-bound forms). Spectra were processed using Topspin (v.4.0.7) and analysed using SPARKY (Research Resource Identifier (RRID): SCR_014228) and CARA (http://cara.nmr.ch/doku.php).

Structural studies in solution

To determine structural details of the NSD2(SET)–IACS-17596 protein–ligand complex, high-resolution multidimensional NMR spectroscopy in solution with various stable isotope labelling strategies and the CNS-based HADDOCK (v.2.4) software⁶⁷, optimized for protein–ligand structure determination, were used as previously described for a similar sized system⁶⁸. In brief, the exact same positions of secondary motifs and fold of NSD2(SET) in the ligand-free and IACS-17596-bound forms were confirmed by Talos-N torsion angle analysis⁶⁶ of the detected backbone chemical shifts, and with the H^N-H^N and H^N-CH₃ (ILV) connectivities in their respective 3D ¹⁵N/¹³C-edited NOESY HSQC spectra of the perdeuterated U-²H,¹³C,¹⁵N-ILV(–¹³CH₃) labelled samples. We failed to crystallize the NSD2(SET)–IACS17596 complex, a result in line with previous reports that the native NSD2(SET) apo form is challenging to crystallize⁶⁹. Therefore, as a starting model for the structural studies, the crystal structure of engineered NSD2(SET) (PDB: 5LSU:A) bearing non-native mutations was adapted to reflect the investigated system. The amino-terminal amino acids up to 983 inclusive were then removed, and the hydrogen atoms and missing side chains were added and optimized in COOT software⁷⁰. The SAM cofactor coordinates were removed in line with the competition experiments, and the three ¹⁰⁷¹LQR¹⁰⁷³ residues were changed to ¹⁰⁷¹DGK¹⁰⁷³, per the NSD2 native ones as in UniProt (O96028). The structural geometry of the ligand was generated and optimized using Avogadro software⁷¹. The CNS structural geometry toppar library files were generated using the full-atom PRODGR algorithm⁷². The residues that exhibited significant amide CSPs after binding to IACS-17596, mainly found in the SAM-binding pocket, were assigned as active and semi-flexible and were used to generate ambiguous interaction restraints. The amide CSPs that were spread across the protein backbone inside the hydrophobic core or belonged to the methyl-containing groups, for which side chains did not demonstrate chemical shift changes of their respective methyl ¹H/¹³C resonances and did not show NOE signals on the ¹³C-edited/filtered NOESY-HSQC (mixing time of 100 ms), like L1151 and V1166, were excluded from defining the ambiguous interaction restraints. The unambiguous interproton intermolecular ligand to protein contacts were identified from ¹³C-edited/filtered NOESY-HSQC (mixing time of 100 ms for the uniformly ¹³C,¹⁵N labelled NSD2(SET)–IACS-17596 complex) and were added to the unambiguous restraints category, all set to a 2.2–6.5 Å distance. The broadened residues beyond the detection on 2D ¹H/¹⁵N correlation TROSY spectra and not displaying any NOE contacts, except its own spin system in the 3D ¹⁵N- and ¹³C-edited NOESY-HSQC spectra, were assigned as fully flexible. A total of 500 structures were submitted for calculation using the HADDOCK (v.2.4) procedure for small ligand binding against NMR-derived restraints. The final ensemble of ten structures showing the lowest combined HADDOCK score was studied and analysed using MolProbity⁷³ and PROCHECK⁷⁴ software. The structural statistics of the determined NSD2(SET)–IACS-17956 structure in solution is provided in Supplementary Table 5. The coordinates and associated experimental datasets have been deposited under the PDB identifier: 9CVD.

Histone extraction and liquid chromatography with tandem MS/MS analysis

Cells cultured with IACS-17596 (or DMSO as control) were prepared for histone extraction. MiaPaCa2 cells were collected after 1, 5 and 9 days of 100 nM IACS-17596 or DMSO treatment. Other cell lines (KP-2, PSN-1, A549 and H1373) were collected after 5 days of 100 nM IACS-17596 or DMSO treatment. For each condition, we performed biological triplicates. The histones were extracted and prepared for chemical derivatization and digestion as previously described^75,76. In brief, the lysine residues from histones were derivatized with propionylation reagent (1:2 reagent to sample ratio) containing acetonitrile and propionic anhydride (3:1), and the solution pH was adjusted to 8.0 using ammonium hydroxide. Propionylation was performed twice and the samples were dried using a speed vac. The derivatized histones were then digested with trypsin at a 1:50 ratio (w/w) in 50 mM ammonium bicarbonate buffer at 37 °C overnight. The N termini of histone peptides were derivatized with propionylation reagent twice and dried using a speed vac. The peptides were desalted with the self-packed C18 stage tip. The purified peptides were then dried and reconstituted in 0.1% formic acid. A liquid chromatography with tandem MS/MS (LC–MS/MS) system consisting of a Vanquish Neo UHPLC coupled to an Orbitrap Exploris 240 (ThermoFisher) was used for peptide analysis. Histone peptide samples were maintained at 7 °C on sample tray for LC. Separation of peptides was carried out on an Easy-Spray PepMap Neo nano-column (2 μm, C18, 75 μm × 150 mm) at room temperature with a mobile phase. The chromatography conditions consisted of a linear gradient from 2 to 32% solvent B (0.1% formic acid in 100% acetonitrile), in solvent A (0.1% formic acid in water) over 48 min and then 42 to 98% solvent B over 12 min at a flow rate of 300 nl min⁻¹. The mass spectrometer was programmed for data-independent acquisition. One acquisition cycle consisted of a full MS scan and 35 data-independent acquisition MS/MS scans of 24 m/z isolation width starting from 295 m/z to reach 1,100 m/z. Typically, full MS scans were acquired in the Orbitrap mass analyzer across 290–1,100 m/z at a resolution of 60,000 in positive profile mode with an automaximum injection time and an automatic gain control (AGC) target of 300%. MS/MS data from higher-energy collisional dissociation (HCD) fragmentation were collected in the Orbitrap. These scans typically used a nominal collision energy (NCE) of 30, an AGC target of 1,000% and a maximum injection time of 60 ms. Histone MS data were analysed using EpiProfile⁷⁷.

Cell culture for RNA-seq, CUT&RUN and NicE-seq

Cells were treated with either IACS-17596 or DMSO at day 0. Medium was refreshed every 2 days. At each collection time point (days 1, 5 and 9), for RNA-seq, 1 million cells per replicate (4 replicates for each time point) were trypsinized and collected. For CUT&RUN, 17 million cells per replicate (2 replicates for each time point) were fixed with 0.1% formaldehyde for 1 min at room temperature, quenched in 125 mM glycine with gentle vortexing for 1 min, washed, resuspended in cryopreservation buffer (growth medium supplemented with 10% DMSO) slowly frozen (–1 °C min⁻¹) and stored at −80 °C until further processing for library preparation. Around 8.5 million cells were recovered per replicate after fixation. For NicE-seq, 2 million cells per replicate (2 replicates for each time point) were crosslinked with 4% formaldehyde for 10 min at room temperature, quenched in 125 mM glycine for 5 min, washed, resuspended in cryopreservation buffer (growth medium supplemented with 10% DMSO), slowly frozen (–1 °C min⁻¹) and stored at −80 °C until further processing for library preparation. Around 1 million cells per replicate were recovered after fixation.

RNA-seq

Total RNA was extracted from MiaPaCa2, KP2 and H1373 cells (with or without treatment with DMSO or IACS-17596) with RNeasy Plus kits (Qiagen) per the supplier’s instructions. Extracted RNA samples were sent to BGI for library preparation and sequencing.

CUT&RUN

CUTANA CUT&RUN (EpiCypher) was performed with lightly fixed MiaPaCa2 cells (with or without treatment with DMSO or IACS-17596) using an automated protocol (autoCUT&RUN) derived from previously described methods^51,78,79. All steps were performed on Tecan Freedom EVO robotics platforms with gentle rocking for incubation steps and magnetic capture for medium exchange–washing steps. In brief, for each reaction, 500,000 cells (at 5 million cells per ml in 20 mM HEPES pH 7.5, 150 mM NaCl, 0.5 mM spermidine and 1× EDTA/EGTA-free complete protease inhibitor (Roche, 11873580001)) were dispensed into individual wells of a 96-well plate, immobilized to concanavalin A beads (EpiCypher, 21–1401) and incubated overnight (4 °C) with the desired antibodies (IgG (EpiCypher 13–0041), H3K4me1 (EpiCypher 13–0057), H3K4me3 (EpiCypher 13–0042), H3K9me3 (ThermoFisher 720093), H3K27me3 (EpiCypher 13–0055), H3K27ac (EpiCypher 13–0045), H3K36me2 (EpiCypher clone: Eab-2527–1B4), H3K36me3 (EpiCypher 13–0058) and NSD2 (EpiCypher 13–0002); all at 0.5 μg per reaction; all histone post-translational modification reagents were validated to fully defined nucleosome standards as previously described^51,78–80). pAG-MNase (EpiCypher, 15–1016) was added and activated with 2 mM CaCl₂ in wash buffer (2 h at 4 °C), and CUT&RUN-enriched DNA was purified using Serapure beads after mixing at 2:1 (bead to DNA) ratio. Recovered DNA was quantified using PicoGreen (ThermoFisher, P11495) and reactions were adjusted to 5 ng DNA (or the entire reaction was used if <5 ng DNA was recovered) before preparing sequencing libraries with a CUTANA CUT&RUN Library Prep kit (EpiCypher, 14–1001) as per the supplier’s instructions. Sequencing (targeting 5–7 million mappable 2 × 50 bp paired-end reads per reaction) was performed on an Illumina NextSeq2000.

NicE-seq

Accessible chromatin was profiled by One-Pot Universal NicE-seq as previously described⁸¹ with minor modifications. In brief, cells (with or without treatment with DMSO or IACS-17596) were crosslinked with 4% formaldehyde and stored at −80 °C. After thawing, nuclei were extracted and 200,000 were used for accessible chromatin labelling and protection (limited Nt.CviPII digestion and repair in the presence of biotinylated and 5-methyl dCTPs). Unprotected genomic DNA was then fragmented with excess Nt.CviPII (37 °C for 2 h), followed by enzyme inactivation (65 °C for 10 min). To capture biotin-labelled DNA, 30 μl magnetic streptavidin beads (NEB S1420S) was added to each reaction and incubated (4 °C for 2 h) in 2 M NaCl. Libraries were prepared and amplified (seven PCR cycles) with a NEBNext Ultra II DNA Library Prep kit (NEB, E7645S), purified on AMPure beads and sequenced (targeting 20–30 million mappable 2 × 50 bp paired-end reads) on an Illumina NextSeq2000.

RNA-seq analysis

RNA-seq raw sequencing reads underwent initial processing to ensure high data quality. Adapter sequences were trimmed and low-quality reads were filtered using CutAdapt with TrimGalore (RRID:SCR_011847). Quality-control metrics for both raw and processed reads were assessed using FastQC (RRID: SCR_014583) and SAMtools (RRID:SCR_002105), respectively. Trimmed reads were aligned to the human reference genome (hg38) using the STAR⁸² aligner, guided by Ensembl gene annotations provided in the GTF format. Post-alignment quality assessment of aligned reads was summarized using multiQC⁸³. Gene-level expression quantification was performed using feature-Counts (RRID:SCR_012919) to count the number of reads mapped to each gene, and transcript per million (TPM) values were computed using salmon, both using the human Ensembl gene GTF annotation. Differential expression analysis between both conditions at each time point of experiments was conducted using DESeq2 (ref. 84). This analysis accounted for biological variability and controlled the false discovery rate. Genes with low counts (<20 reads in more than half of all samples) were excluded from further analysis. Analysis was limited to ‘protein-coding’ genes (Human Genome Organization Gene Nomenclature Committee, April 2024 version), and genes were considered significantly dysregulated if they met a false discovery rate adjusted P value cut-off of 0.05 and a fold change cut-off of 1.5. GSEA (done on all expressed genes, ranked on the basis of decreasing log₂[fold change]) was performed to identify biological pathways enriched with dysregulated genes. Enrichment analysis was conducted using the fgsea⁸⁵ package in R, with publicly available and custom gene signatures. Enrichment results were further analysed and illustrated using the clusterProfiler⁸⁶ and ggplot2 R packages⁸⁷, with NES plots showcasing pathways of interest (for which all statistical tests were performed and results were considered significant if P < 0.05 after adjusting for multiple testing).

Custom gene sets

We defined PRC2 target genes according to the presence of a peak of H3K27me3 at their promoters. Narrow and broad peaks were called using macs2 (RRID: SCR_013291) with default parameters in DMSO day 1 data. Those peaks were intersected with the TSS annotation to identify target genes. Genes were then ranked according to the H3K27me3 read counts within 1 kb of the promoter region flanking the TSS. As many PRC2 target genes are transcriptionally silent, to obtain a list for further expression analysis, we intersected this list with the output of DESeq2 (ref. 84) and retained the top 500 PRC2 target genes with detectable expression. PRC2 target neighbours were defined as genes within 400 kb of the PRC2 targets but excluding the PRC2 targets themselves. Enhancer-dependent genes were defined using the Genhancer⁸⁸ database. Genes ranked first in the GeneHancer.gff file attribute ‘connected_gene’ were extracted, and their enhancers were counted. For genes that are not included in the GeneHancer database, the enhancer count was considered zero. Genes were ranked by the number of enhancers, and the top 1,000 were selected as enhancer-dependent.

Control gene sets

To compare characteristics of target gene sets with appropriate controls, for each gene set we created an expression-matched set of controls. TPMs in vehicle-treat samples (n = 4) from day 1 were used as the baseline for expression levels. Genes with low counts were excluded, following the same criteria as for RNA-seq analysis. For each gene in target gene sets, such as GSEA-identified signatures and PRC2 target genes, the Euclidean distance was calculated across the TPM matrix of day 1 vehicle-treated samples. The five nearest neighbours from the complement of the gene list were retained, which resulted in a fivefold replicated control gene set. To compare the expression of target genes with selected control genes, the TPM values of the five control genes for each target gene were averaged. A Wilcoxon signed-rank paired test was then performed to assess differences of TPMs between the target gene set and the averaged control gene set.

NicE-seq analysis

NicE-seq raw reads underwent quality control and trimming using CutAdapt⁸⁹ with TrimGalore (RRID:SCR_011847). Sequences with low-quality bases (Phred score < 20) and adapter contamination were removed. The resulting trimmed reads were aligned to the human genome assembly (hg38) reference genome using Bowtie2 (ref. 90) with default parameters. After alignment, non-uniquely mapping reads were removed and remaining reads were further filtered with samclip (https://github.com/tseemann/samclip) to ensure high mapping specificity. The identification of peaks indicative of chromatin accessibility was performed using MACS2 (ref. 91) with parameters ‘-extsize 200 -shift –100 -nomodel’. This approach identifies peaks based on the enriched fragment size and shifts the pileup to the peak summit for optimal peak calling accuracy. To ensure robustness of identified peaks across experimental replicates, the irreproducible discovery rate (IDR) was computed using the IDR framework. Peaks with IDRs greater than 0.05 across every pairwise comparison of replicates were excluded from downstream analysis. The resulting peak sets were further refined by removing peaks that overlapped with regions identified in the ENCODE blacklist⁹². This step minimizes artefacts associated with genomic regions prone to technical biases or artefacts. NicE-seq principal component analysis (PCA) plots were generated by running featureCounts⁹³ on a union of peaks, running different peak analysis with DESeq2 (ref. 94) and filtering for the top 500 peaks with the highest variance across samples. Normalized coverage tracks were generated using deepTools⁹⁵ ‘bamCoverage with parameters ‘-b $BAM -o $OUTPUT.bigWig --normalizeUsing CPM --centerReads -e 200 --minMappingQuality 5 -bs 200’. This step enables visualization of read coverage across the genome, normalized to counts per million (CPM), with reads centred around the peak summit for enhanced resolution.

CUT&RUN analysis

CUT&RUN raw sequencing reads underwent initial quality assessment and adaptor sequence trimming using CutAdapt with TrimGalore (RRID: SCR_011847). Reads containing low-quality bases (Phred score < 20) and adapter sequences were removed to generate trimmed reads that were then aligned to a combined reference genome consisting of the human genome assembly (hg38) and the E. coli genome assembly (ec57) using Bowtie2 (RRID:SCR_016368). Alignments with low mapping quality (MAPQ < 3) were filtered out using SAMtools⁹⁶ to ensure high-quality alignments for downstream analysis. To further refine data quality, anomalous signals, such as contaminated mapped sequences, were filtered using SAMtools in conjunction with samclip. This step was performed on all samples to enhance the reliability of subsequent analyses. Resulting mapped reads were further processed to quantify CUT&RUN signals by binning raw tag counts into genomic windows of different sizes (1 kb and 10 kb) using BEDTools⁹⁷. Normalization of CUT&RUN signals was performed using deepTools (bamCoverage) (RRID: SCR_016366). For each sample replicate, normalization was achieved by either (1) dividing by the total number of aligned reads in millions (CPM normalization), in cases when absolute quantification by MS (described below) was not possible or (2) MS normalization, involving scaling the CPM signal by a factor to ensure that the mean across all genomic bins is equal to MS fraction (averaged across replicates) corresponding to the profiled histone modification. The MS-normalized signal approximates the local absolute level of each modification and enables quantitative comparisons across conditions. We also applied E. coli spike-in normalization to ensure concordance with the MS approach. In brief, scaling factors for each sample were determined by the number of reads mapped to the exogenous genome (STEC 1686) and introduced a constant of 1,000,000 to achieve reference-adjusted reads per million normalization⁹⁸. PCA and hierarchical clustering were performed using the log₂[depth-normalized] signals (with the addition of a pseudocount (1 × 10⁻¹⁵) to avoid division by 0) derived from 10-kb genomic bins. These analyses provided insights into the similarity and clustering of samples based on their CUT&RUN profiles. Correlation plots between CUT&RUN signals were generated by merging replicates per condition and using the log₂[depth-normalized] signals scaled per million derived from 10-kb genomic bins. To generate merged coverage tracks across replicates, normalized signals were integrated using bigwigCompare from deepTools (RRID: SCR_016366). The merging process was conducted stepwise with the parameter setting ‘--operation mean -bs 200’, which ensured that combined signals accurately represented the cumulative data from all replicates. Genic and intergenic binned analysis plots were generated by merging replicates per condition and using log₂[depth-normalized] and MS-normalized signals (with the addition of a pseudocount (1 × 10⁻¹⁵) to avoid division by 0) derived from 10-kb genomic bins. Genic regions were designated as the union of any genomic intervals having the ‘gene’ annotations in Ensembl⁹⁹ and intergenic regions were defined as the complement of genic ones (that is, genomic regions not overlapping genic regions). Bins overlapping genic regions but not intergenic regions were defined as genic bins and vice versa for bins overlapping intergenic regions. These analyses provided insights into region-specific genome-wide changes occurring for a given broad histone mark.

Fraction of reads in peaks calculations

Fraction of reads in peaks scores for NicE-seq and CUT&RUN samples were calculated to quantify the proportion of sequencing reads enriched in peak regions. First, the total number of aligned reads for each sample was determined using SAMtools. Then, reads overlapping peak intervals were counted using BEDTools. The fraction of reads in peaks score was computed as the ratio of reads intersecting peaks to the total number of aligned reads.

Signals on promoters of target versus control gene sets

The promoters of each transcript in the target and control gene sets were identified using the R package GenomicRanges based on a region extending 1,000 bp upstream of the TSSs. Annotations were obtained from the TxDb.Hsapiens.UCSC.hg38.knownGene package, and the identified promoter regions were subsequently converted to BED format using the rtracklayer package. The ComputeMatrix command from deepTools was used to calculate the average CUT&RUN (H3K27me3 and H3K36me2) and NicE-seq signals across promoter regions. To evaluate differences between signal profiles, the central bins were extracted from the ComputeMatrix output (CUT&RUN: 25 bins = 5 kb; NicE-seq: 20 bins = 1 kb). Significance was determined using the two-sided Wilcoxon signed-rank paired test.

Visualization

We extracted protein-coding genes from the Ensembl gene annotation file and defined TSS and TES coordinates based on the most upstream start and downstream end of their transcripts, respectively. Only genes with a TPM expression greater than zero were included to avoid noise from potentially fully heterochromatized genes. Intergenic regions were defined as regions larger than 20 kb that do not overlap with genes. The abundance of various histone marks and NSD2 on gene bodies and intergenic regions was visualized using deepTools ComputeMatrix and plotHeatmap functions (RRID: SCR_016366). Coverage and viewer tracks were visualized using pyGenomeTracks¹⁰⁰.

Analysis of datasets from The Cancer Genome Atlas

The Cancer Genome Atlas database with available gene expression data for human cancer and normal tissue samples were analysed using cBioPortal (v.5.3.13)¹⁰¹ to evaluate NSD2 mRNA expression in indicated cancer types. Basal-normalized transcript expression data (z scores) used for this analysis were RNA Seq V2 RSEM.

Animal models

Kras^cKI-G12C, Trp53^loxP/loxP, Rosa26^{LSL-NSD2(E1099K)} and Ptf1a^cre mice have been previously described^4,39,102,103. Mice were maintained on a mixed C57BL/6;129S1 strain background and we systematically used littermates as controls in all the experiments. Both male and female animals were used in the experiments, and no sex differences were noted. Immunocompromised NSG mice (NOD.SCID-IL2Rg^−/−, Jackson Laboratories, strain 005557) were used for xenograft studies. All NSG xenograft experiments were performed on 6–10-week-old female animals. In all experiments, animals were numbered, and experiments were conducted in a blinded fashion. After data collection, genotypes were revealed, and animals were assigned to groups for analysis. For treatment experiments, mice were randomized. None of the mice with the appropriate genotype were excluded from this study or used in any other experiments. All mice were co-housed with littermates (2–5 per cage) in a pathogen-free facility with standard controlled temperature of 22 °C, with a humidity of 30–70% and a light cycle of 12 h on–12 h off set from 7:00 to 19:00 and with unrestricted access to standard food and water under the supervision of veterinarians in an AAALAC-accredited animal facility at the University of Texas MD Anderson Cancer Center (MDACC). Mouse handling and care followed the NIH Guide for Care and Use of Laboratory Animals. All animal procedures followed the guidelines of and were approved by the MDACC Institutional Animal Care and Use Committee (IACUC protocol 00001636, principal investigator (PI): P.K.M.). Tumour size was measured using a digital caliper and tumour volume was calculated using the formula: volume = (width)² × length/2, where length represents the largest tumour diameter and width represents the perpendicular tumour diameter. The end point was defined as the time at which a progressively growing tumour reached 20 mm in its longest dimension as approved by the MDACC IACUC protocol (00001636, PI: P.K.M.). In no experiments was this limit exceeded.

Metabolic stability (ADME) assays in vitro

Microsomal stability.

Microsomal stability assays were conducted using a Beckmann Biomek FXp laboratory automation system. The liver microsomal incubation mixture consisted of liver microsomes (0.5 mg microsomal protein per ml), the test compound (1 μM), MgCl₂ (3 mM) and EDTA (1 mM) in potassium phosphate buffer (100 mM, pH 7.4). Midazolam and ketanserin were used as the assay control substrates. The reaction was initiated with the addition of a NADPH regeneration solution (1.3 mM NADPH) and maintained at 37 °C with shaking. At 5 time points ranging from 0 to 45 min, aliquots (50 μl) were removed and quenched with acetonitrile (100 μl) containing an internal standard (imipramine). After vortexing and centrifugation, the samples were analysed by LC–MS/MS. Calculation of the in vitro half-lives and clearance followed literature guidelines¹⁰⁴.

Hepatocyte stability.

Hepatocytes were purchased from Bioreclamation IVT, Xenotech or RILD. Stock solutions were prepared at 10 mM in DMSO for the test compound. Aliquots of the stock solutions were diluted to 200 μM with DMSO and then further diluted to 2 μM with Krebs−Henseleit buffer. The procedure was as follows: count hepatocytes and then dilute the cell suspensions to the appropriate density (viable cell density = 2 × 10⁶ cells per ml). Add 50 μl prewarmed (37 °C) 2 μM test compound to the wells designated for different time points. For 0 min, add 100 μl acetonitrile containing internal standard to the wells followed by 50 μl hepatocyte solution and then seal the wells. Add 50 μl prewarmed hepatocyte solution to the wells designated for 15, 30, 60 and 120 min and start timing. Place the assay plate in an incubator at 37 °C. At 15, 30, 60 and 120 min, add 100 μl acetonitrile to the wells, then seal the wells. After quenching, sonicate the plate for 5 min and then centrifuge at 5,594g for 15 min (Thermo Multifuge × 3R). Transfer 50 μl of the supernatant from each well into a 96-well sample plate containing 120 μl ultrapure water for LC–MS analysis. The peak area response ratio to internal standard of the compounds at 15, 30, 60 and 120 min was compared to the peak area response ratio at 0 min to determine the per cent of the test compound remaining at each time point. Half-lives were calculated using Excel software, fitting to a single-phase exponential decay equation.

Plasma protein binding.

Plasma protein binding assays were conducted using a rapid equilibrium dialysis device (ThermoFisher). Warfarin and metoprolol were used as the control substrates. In the receiver side, 350 μl PBS (pH 7.4, Gibco) was added. In the donor side, 200 μl of plasma (Bioreclamation IVT) spiked with the test compound (5 μM) was added. The same plasma or test compound solution (50 μl) was also used for the recovery sample. The plate was covered with Immunoware sealing tape and was incubated at 37 °C with shaking at 100 r.p.m. for 5 h. After the incubation step, both the receiver and donor sides were sampled (50 μl) and matched with the same volume of matrix from the other side. The recovery, donor and receiver samples were extracted with 300 μl cold CAN containing imipramine as the internal standard. After vortexing and centrifugation, the supernatant (150 μl) was subjected to LC–MS quantitation. Plasma protein binding (% bound) was calculated as % bound = 100 × ([donor] − [receiver])/[donor].

In vivo pharmacokinetics

All experiments were conducted in accordance with the animal welfare procedures and were approved by the MDACC Institutional Animal Care and Use Committee (IACUC protocol 00001636, PI: P.K.M.). Male C57BL/6 mice, weighing 20−30 g were used for studies. Food and water were available to all the animals ad libitum. NSD2i compounds were intraperitoneally administered at 100 mg kg⁻¹. Blood samples were collected from all the animals at pre-dose and at 0.083, 0.25, 0.5, 1, 2, 4, 8 and 24 h after dosing into tubes containing the anticoagulant K₂EDTA (3 animals per time point with 3 time points collected per animal). Plasma was separated from the blood by centrifugation at 4 °C and stored at −70 °C until analysis. NSD2i concentrations in plasma were quantified by LC–MS/MS.

Toxicity evaluation

All experiments were conducted in accordance with the animal welfare procedures and were approved by the MDACC Institutional Animal Care and Use Committee (IACUC protocol 00001636, PI: P.K.M.). Female and male wild-type C57BL/6 mice aged 8 weeks were administered IACS-17817 (100 mg kg⁻¹ once per day, intraperitoneally) or vehicle (40% (2-hydroxypropyl)-β-cyclodextrin (Sigma-Aldrich, 778966)) for 2 weeks. At the time of necropsy, whole-blood serum and EDTA samples were collected for clinical chemistry and complete blood count analysis, respectively (MDACC Department of Veterinary Medicine and Surgery). Following macroscopic examination, selected organs were collected and fixed in 10% neutral-buffered formalin for 24 h. Tissue samples were embedded in paraffin, sectioned, stained with H&E and examined by microscopy. Markers of cell proliferation (Ki67⁺) and cell death (cleaved caspase-3⁺) were assessed by immunohistochemistry and quantified.

PDAC mouse model and treatment studies

To evaluate the effects of NSD2 hyperactivation on the development and progression of PDAC, we used Kras^cKI-G12C/+, Trp53^loxP/loxP, Rosa26^{LSL-NSD2(E1099K)} and Ptf1a^cre (K^CPC;Nsd2^E1099K) and Kras^cKI-G12C/+, Trp53^loxP/loxP, Ptf1a^cre (K^CPC) animals, which develop aggressive disease. Mice were followed for signs of disease progression. At the indicated time point (4 weeks of age) or the humane end point, pancreatic tumour tissues were processed for biochemical, histological and immunohistochemical analyses. Histopathological analysis was conducted on de-identified slides. To assess the therapeutic efficacy of NSD2i treatment and combination therapy, we used K^CPC pancreatic cancer models. At 7 weeks of age, tumour-bearing mice were randomized and assigned into treatment groups. Mice were treated as indicated with IACS-17817 (100 mg kg⁻¹ once per day, intraperitoneally, Wuxi AppTec) and/or sotorasib (10 mg kg⁻¹ once per day, intraperitoneally, Advanced ChemBlocks) in vehicle (40% (2-hydroxypropyl)-β-cyclodextrin (Sigma-Aldrich, 778966)). Control and monotherapy animals underwent the same procedure but received vehicle treatment. The survival end point was determined by overall health criteria scoring. Mouse health status and weight were checked daily.

LUAD mouse models and treatment studies

To evaluate the effects of NSD2 hyperactivation on the development and progression of LUAD, we used Kras^cKI-G12C, Trp53^loxP/loxP, Rosa26^{LSL-NSD2(E1099K)} (K^CP;Nsd2^E1099K) and Kras^{cKI-G12C/G12C}, Trp53^loxP/loxP (K^C/CP) animals. To generate tumours in the lungs of K^CP;Nsd2^E1099K and control K^C/CP mutant mice, we used Ad-Cre. In brief, 8-week-old mice were anaesthetized by continuous gaseous infusion of 2% isoflurane for at least 10 min using a veterinary anaesthesia system. Virus was delivered to the lungs by intratracheal intubation. Before administration, virus was precipitated with calcium phosphate to improve the delivery of Cre by increasing the efficiency of viral infection of the lung epithelium. Mice were treated with one dose of 5 × 10⁶ p.f.u. of Ad-Cre. Mice were analysed for tumour formation and progression at the indicated time points after viral infection. To assess the therapeutic efficacy of NSD2i treatment and combination therapy, we used the K^CP lung cancer mouse model. At 9 weeks after Ad-Cre delivery, tumour-bearing mice were randomized and assigned to treatment groups. Mice were treated as indicated with IACS-17817 (100 mg kg⁻¹ once per day, intraperitoneally, Wuxi AppTec) and/or sotorasib (10 mg kg⁻¹ once per day, intraperitoneally, Advanced ChemBlocks) in vehicle (40% (2-hydroxypropyl)-β-cyclodextrin (Sigma-Aldrich, 778966)). Control and monotherapy animals underwent the same procedure but received vehicle treatment. The survival end point was determined by overall health criteria scoring. Mouse health status and weight were checked daily.

Xenograft models and treatment studies

PDXs were obtained from the NCI Patient-Derived Models Repository (PDMR), NCI-Frederick, Frederick National Laboratory for Cancer Research (specimen ID: LG0567-F671-PDC and 323965–272-R-J2-PDC). The PDAC PDX sample was obtained from a de-identified patient (70-year-old woman) with histologically confirmed grade 3 poorly differentiated PDAC, treatment naive and with the following mutations: KRAS^G12C, TP53^R282W, CDKN2A^L16Pfs*9, ARID1A^S958Lfs*10, FLT3^V194M and PTPRS^R1821*. The LUAD PDX sample was obtained from a de-identified patient (57-year-old man) with histologically confirmed grade 3 poorly differentiated LUAD, treatment naive and with the following mutations: KRAS^G12C, TP53^R273C, KMT2C^{M3463_Q3464delinsI*}, MSH6^K1358Dfs*, TET1^E1640* and MUTYH^X298_splice. PDX tumour specimens were collected after written patient consent and in accordance with the Institutional Review Board-approved protocols of the MDACC (PA19–0435, PI: P.K.M.). For NSD2i therapy studies, PDX specimens were grafted subcutaneously to NSG mice. When tumours became palpable, they were measured by caliper to monitor growth kinetics. Mice were treated as indicated with the NSD2i compounds (100 mg kg⁻¹ once per day, intraperitoneally, Wuxi AppTec) and sotorasib (10 mg kg⁻¹ once per day, intraperitoneally, Advanced ChemBlocks) in vehicle (40% (2-hydroxypropyl)-β-cyclodextrin (Sigma-Aldrich, 778966)). Control and monotherapy animals underwent the same procedure but received vehicle treatment.

Histology and immunohistochemistry

Tissue specimens were fixed in 4% buffered formalin for 24 h and stored in 70% ethanol until paraffin embedding. H&E staining and immunostainings were performed on 3-μm-thick tissue sections. Tissue microarray patient specimens were collected after written patient consent and in accordance with the Institutional Review Board-approved protocols of the MDACC (PA19–0435, PI: P.K.M.). The following antibodies were used (at the indicated dilutions): cleaved caspase-3 (CST, 9664; 1:200) and Ki67 (550609 BD Bioscience; 1:1,000). Immunohistochemistry was performed on formalin-fixed, paraffin-embedded tissue sections using a biotin-avidin HRP conjugate method (Vectastain ABC-HRP kit, PK4000) as previously described¹⁰⁵. Sections were developed with DAB and counterstained with haematoxylin. Pictures were taken using a PreciPoint M8 microscope equipped with PointView software and quantified using ImageJ software (v.1.53k, RRID: SCR_003070) and QuPath (v.0.5.1, RRID: SCR_018257).

Single-cell isolation from tumour samples

Freshly microdissected lung tumours from K^C/CP mice undergoing the indicated treatment were processed to prepare single-cell suspensions. Specifically, tumours were cut into small pieces and digested with the enzyme mix provided in a Tumour Dissociation kit (Miltenyi Biotec). Samples were dissociated into single-cell suspensions with a gentleMACS dissociator (Miltenyi Biotec), followed by density-gradient centrifugation using Debris Removal solution (Miltenyi Biotec). Cells were filtered through 70 μm strainers, centrifuged and resuspended in ice-cold ACK lysis buffer (ThermoFisher) for red blood cell lysis. Samples were washed, counted and subjected to scRNA-seq.

scRNA-seq

Cells were processed for scRNA-seq at The University of Texas MD Anderson Cancer Center Advanced Technology Genomics Core. scRNA gene expression libraries were prepared using a 10x Genomics Chromium Next GEM Single Cell 5′ Reagent Kit v.2. Libraries were sequenced using NovaSeq 6000 S4 flow cells (Illumina). scRNA-seq data were processed as previously reported¹⁰⁶. In brief, gene expression reads were aligned to the mm10 genome (mouse) using the 10x Genomics Cell Ranger¹⁰⁷. Filtered feature-barcode matrices were further analysed in R using Seurat (v.4.4.0)¹⁰⁸. In general, genes detected in <3 cells and cells expressing <200 RNA features (genes) were filtered out. Low-quality cells containing a high percentage of mitochondrial reads were also excluded from subsequent analyses. Cells identified as doublets by DoubletFinder (v.2.04)¹⁰⁹ were removed. Cells passing quality control were subjected to the standard Seurat workflow. After PCA, uniform manifold approximation and projection (UMAP) reduction and clustering was performed to identify and visualize cell clusters. Clusters co-expressing multiple lineage markers were further excluded as doublets. The pipeline was rerun with the remaining cells for downstream analysis. DEGs between two groups were identified using the FindMarkers function in Seurat. GSEA was performed using clusterProfiler (v.4.10.0)¹¹⁰ on curated gene sets from the Molecular Signatures Database. Additional gene signatures were retrieved from previous studies, specifically: ‘Xue lung KRAS induced’³⁶, ‘pancreatic neoplasia up/activated’⁵³, ‘alveolar signature (MP31)’¹¹¹ and ‘tumour fitness score’⁵⁸. NES and adjusted P values were calculated by two-sided permutation test as previously described¹¹⁰. For comparative analysis, malignant tumour cells were distinguished from non-malignant normal cells using inferCNV¹¹², with normal mouse lung alveolar cells⁸⁴ serving as the reference control. The Bhattacharyya distance was used to assess transcriptomic variance between tumour and normal epithelial cells across different treatment groups, with minor modifications to a previously described approach¹¹³. In brief, cells were embedded into PCA space, retaining the top 25 principal components for calculations. For each comparison, 100 cells were randomly sampled from each group, repeated 100 times, and the Bhattacharyya distance was computed. P values by two-sided Mann–Whitney–Wilcoxon unpaired test lower than the calculation limit were reported as <2.2 × 10⁻¹⁶. The leading-edge genes, defined as the subset of genes that contribute most to the enrichment result¹¹⁴, were extracted from corresponding GSEA results comparing vehicle-treated and NSD2i-treated tumour cells. To compare gene expression values across sample types, the normalized gene values were z score normalized.

Tumour cell fitness scoring and reference mapping

High-fitness-associated genes derived from a mouse model of Kras; Trp53 (KP)-driven lung adenocarcinoma were retrieved from previous study⁵⁸. Single-cell gene set scoring was performed using UCell (v.2.6.2.)¹¹⁵. For KP-Tracer reference mapping, the reference was downloaded from Zenodo (https://doi.org/10.5281/zenodo.5847462), and epithelial cells from different treatment groups were used as the query. Data transfer from the reference to the query was performed in Seurat. For UMAP visualization, cells from the query dataset were mapped to the reference using k-nearest neighbour method based on their similarity. Cell density was plotted using contourPlot from the R package scDataviz (v.1.1.2, https://github.com/kevinblighe/scDataviz).

Compound synthesis

For detailed description of synthesis and characterization of compounds see Supplementary Methods.

Quantification and statistical analysis

Please refer to the figure legends or the experimental details for description of sample size (n) and statistical details. All values for n are for individual mice or individual samples. Sample sizes were chosen on the basis of previous experience with given experiments. Cell culture assays were performed in triplicate and in two independent experiments unless stated otherwise. Data are expressed as the mean ± s.e.m. Differences were statistically analysed as indicated by log-rank, unpaired two-sided t-tests, Wilcoxon signed-rank tests, Mann–Whitney–Wilcoxon tests, Wald tests or two-way ANOVA with Tukey testing for multiple comparisons.

Statistics and reproducibility

All statistics used in this study are described when used in the appropriate section. For all the western blots, autoradiography and cell culture assay results, the reported data were reproduced three independent times (or more as indicated), and representative data are shown. No statistical methods were used to predetermine sample sizes. All western blot analyses and micrographs, unless otherwise stated, are from three independent biological replicates for each time point and condition (Figs. 1c,h, 3b and 5f,i and Extended Data Figs. 1e,f,j, 2d–h,j, 4a and 7a,e). For source data see Supplementary Fig. 1. CUT&RUN and NicE-seq data are from two independent biological replicates for each time point and condition. Bulk RNA-seq data are from four (MiaPaCa2 cells) and three (KP2 and H1373 cell lines) independent biological replicates for each time point and condition. qMS are all from three independent biological replicates for each time point and condition.

Extended Data

Extended Data Fig. 1 |. NSD2 is overexpressed in solid tumours and accelerates KRAS-driven tumorigenesis in vivo.

a, PRC2 methylation activity is blocked by H3K36me2. In vitro methylation assays as in (Fig. 2b) with recombinant PRC2 complex (left) and NSD2 (right) on unmodified or modified rNuc (nucleosomes) substrate as indicated. Data are means ± s.e.m. from three independent experiments. P values determined by two-way ANOVA with Tukey’s testing for multiple comparisons. b, NSD2 expression is elevated across diverse cancer types relative to cognate normal samples (TCGA data). c, Schematic of generation of Kras^G12C/+;p53^lox/lox (K^CP) mouse LUAD model expressing catalytically hyperactive Nsd2^E1099K allele (K^CP;Nsd2^E1099K). d, Experimental design to assess effects of NSD2^E1099K expression on LUAD pathogenesis in K^CP model. Lung-specific recombination of mutant alleles is induced by intratracheal lavage of adenovirus expressing Cre-recombinase (Ad-Cre). e, NSD2 expression and levels of H3K36me2 are elevated in K^CP mouse model of LUAD compared to normal lung. Western blot analysis with indicated antibodies of two representative lysates of tissue biopsies from normal lung and LUAD from K^CP mice. Vinculin and total H3 shown as loading controls. Arrowhead: NSD2. f, Western analysis with indicated antibodies of representative tissue biopsy lysates from K^CP and K^CP;Nsd2^E1099K LUAD samples at 7 weeks post-tumour induction. Vinculin and total H3 are loading controls. Arrowhead: NSD2. g, Representative macro pathology, HE staining, and IHC staining for markers of proliferation (Ki67⁺) and cell death (cleaved Caspase3⁺) of lung tumours from K^CP and K^CP;Nsd2^E1099K mutant mice at 7 weeks after tumour induction and normal lungs from wild-type mice (representative of n = 7/group), scale bars: 5 mm for macro pathology, 3 mm for HE; 50 μm for HE zoom and IHC. Arrowheads indicate cleaved Caspase3 positive cells. h, Quantification of tumour burden, tumour number, proliferation (Ki67⁺) and cell death (cleaved Caspase3⁺) in K^CP and K^CP;Nsd2^E1099K mutant mice as in (g). P values determined by two-tailed unpaired t-test. Data are means ± s.e.m. of n = 7 mice per group. Boxes: 25^th to 75^th percentile, whiskers: min. to max., centre line: median. i, Kaplan-Meier survival curves of K^CP control (n = 12, median survival 160 days) and K^CP;Nsd2^E1099K mutant mice (n = 7, median survival 64 days). P values determined by a log-rank test. Data are means ± s.e.m. of indicated number of biological replicates. j, NSD2 expression and levels of H3K36me2 are elevated in K^CPC mouse model of PDAC compared to normal pancreas. Western blot analysis with indicated antibodies of two representative lysates of tissue biopsies from normal pancreas and PDAC from K^CPC mice. Vinculin and total H3 shown as loading controls. Arrowhead: NSD2. k, Experimental design to assess effects of NSD2^E1099K expression on PDAC pathogenesis in K^CPC model. l, IHC staining for proliferation (Ki67⁺) and cell death (cleaved Caspase3⁺) of pancreatic tumours from K^CPC and K^CPC;Nsd2^E1099K mutant mice at 4 weeks of age (representative of n = 5/group), scale bars: 50 μM. Arrowheads indicate positive cells. m, Quantification of proliferation (Ki67⁺) and cell death (cleaved Caspase3⁺) in K^CPC and K^CPC;Nsd2^E1099K pancreas samples as in (l). P values determined by two-tailed unpaired t-test. Data are means ± s.e.m. of n = 5 mice per group. Boxes: 25^th to 75^th percentile, whiskers: min. to max., centre line: median.

Extended Data Fig. 2 |. Characterization of NSD2i activity, potency, and selectivity.

a, Potency of NSD2i compounds against a KMT panel. Dose-response curve of IACS-17596 (left) and IACS-17817 (right) inhibition of in vitro methylation activity for the indicated KMTs. IC₅₀ values are provided in Fig. 1g. Relative activity for each KMT is normalized to control conditions. Data are means ± s.e.m. from n = 9 (NSD1–3) and n = 6 (other KMTs) independent experiments. b, NSD2i inhibits in vitro methylation activity of gain-of-function NSD2 variants NSD2_SET-E1099K and NSD2_SET-T1150A. In vitro methylation assays as in Fig. 2f. Error bars represent s.e.m. from three independent experiments. P value determined by two-way ANOVA for multiple comparisons. c, NSD2i is not functionally active against a panel of human kinases. Kinome profiling of IACS17596 and IACS17817 at 1 μM against 97 human kinases (see Supplementary Table 1). d, NSD2i treatment depletes H3K36me2 in cells. Western analysis with the indicated antibodies of WCEs from MiaPaCa2 cells treated for 48 h with IACS-17596 (NSD2i; concentrations as indicated). H3 and tubulin are loading controls. e, Western analysis of WCEs with the indicated antibodies in MiaPaCa2 cells treated for 48 h with IACS-17817 at indicated concentration. H3 is shown as loading control. f, Evaluation of the level of H3K36me2 at 12 h and 24 h after treatment of IACS17596 (left) and IACS17817 (right). Western analysis of WCEs with the indicated antibodies in MiaPaCa2 cells treated for 12 h and 24 h with the indicated compound at 100 nM. H3 is shown as loading control, g, NSD2i treatment depletes H3K36me2 in mouse cancer cell lines. Western analysis of WCEs with indicated antibodies in KP LUAD and KPC PDAC mouse cancer cell lines and treated for 48 h with 100 nM of IACS-17596 and IACS-17817 as indicated. Arrowhead: NSD2; asterisks: non-specific band. Tubulin and total H3 are shown as loading controls. h, NSD2i treatment does not further deplete H3K36me2 in clonal NSD2 KO mouse cancer cell lines. Western analysis of WCEs with the indicated antibodies in 4T1 and B16F10 control or clonal NSD2 knockout cells as indicated and treated with 1 μM NSD2i or control (DMSO) for 72 h. H3 and tubulin are shown as loading control. Arrowhead: NSD2; asterisks: non-specific band. i, NSD2i inhibits colony formation of indicated human PDAC cancer cell lines. Cell lines were treated with NSD2i 1 μM for 2 weeks or control (DMSO). Quantification of colonies from 3 independent experiments. P values determined by two-tailed unpaired t-test. j, Western analysis using the indicated antibodies of WCEs from the indicated NSCLC and melanoma cell lines ±100 nM IACS-17596 for 48 h as indicated. k, Cellular viability dose-response curves (left) and IC₅₀ values shown (right) for the cell panel as in (j) treated with IACS-17596 at the indicated concentrations. Cell viability was determined after 7 days of treatment. Data are means ± s.e.m. from three independent experiments.

Extended Data Fig. 3 |. Structural and biophysical analyses of IACS-17596 and IACS-17817.

a, IACS-17596 is SAM competitive. Close-up of the methyl group proton signal from the SAM cofactor in free form. The free SAM signal doubles when the 17596 ligand is added. The samples were measured on 800 MHz Bruker Avance NEO NMR spectrometer at 30 °C. The concentrations of NSD2_SET, SAM, and 17596 were 150 μM in 50 mM Tris pH=7.5, 150 mM NaCl, 1 mM TCEP. b, NMR monitored binding of IACS-17596 (left) and IACS-17817 (right) ligand to NSD2_SET. The aliphatic methyl region from the 2D 1H-13C SOFAST-methyl-TROSY NMR spectra of uniformly double labeled 13 C,15 N apo NSD2_SET (red) and in complex with either IACS-17596 (blue, top left) or IACS-17817 (blue, top right). Selected methyl groups are assigned and labelled. The samples were measured on 800 MHz Bruker Avance NEO NMR spectrometer at 30 °C. The concentration of NSD2_SET was 200 μM in 50 mM Tris pH=7.5, 150 mM NaCl, 1 mM TCEP buffer. The 2D 1H-15N HSQC-TROSY spectra of perdeuterated U-2H,13 C,15N-ILV (–13CH3) apo NSD2_SET (red) and in complex saturated with either IACS-17596 (blue, bottom left) or IACS-17817 (blue, bottom right). Selected residues experiencing chemical shift perturbations are labelled. c, Similar binding mode for IACS-17596 and IACS-17817 interaction with NSD2_SET. The ligand induced amide chemical shift perturbation (CSP_N/H) plots of IACS-17596 (left) and IACS-17817 (right). The CSPs were determined from 2D 1H-15N HSQC-TROSY spectra of perdeuterated U-2H,13 C,15 N ligand-free and ligand bound NSD2_SET forms. The grey squares denote the prolines and not detected residues in both ligand free and ligand bound forms of NSD2_SET, while cyan are the residues that broaden beyond the detection after NSD2_SET binds to ligands. d, Observed backbone amide CSPs, as in (c), induced by IACS-17817 (left panel) and IACS-17596 (right panel) are mapped on the 3D NSD2_SET solution structure (PDB 9CDV; for clarity the ligand coordinates are removed). e, Multiple-sequence alignment of human NSD1, NSD2 and NSD3 SET domains derived from Uniprot_Q96L73, Uniprot_O96028, and Uniprot_Q9BZ95, respectively. Red colour represents fully conserved amino acid residue, pink colour represents highly conserved amino acid residue. Residues corresponding to L1150 and V1166 of NSD2 are highlighted in blue.

Extended Data Fig. 4 |. NSD2i regulation of histone modification abundance and chromatin distribution.

a, Western analysis WCEs expansion of Fig. 3b with the indicated antibodies. Note H3K36me2, NSD2, H3 and Tubulin blots are the same as those used in Fig. 3b. b, Heatmap profile of histone modifications assessed by qMS as in Fig. 3c in the indicated PDAC and LUAD cell lines treated with DMSO or 100 nM IACS-17596 for five days. Data from 3 independent biological replicates. Log₂ fold change values (left) were normalized to the relative abundance in DMSO-treated control group for each histone modification. Arrowheads indicate H3.1K36me2 and H3.3K36me2. c, Representative tracks (as in Fig. 3d) for each biological replicate of MS-normalized CUT&RUN signals from the qMS histone analysis, and read depth-normalized H3K4me3, NSD2, and NicE-seq signals on region of chromosome 2, D1 control samples (as in Fig. 3d). Mb: megabase. Genes present in region are indicated at the bottom. d, Representative tracks (as in Fig. 3f) for each biological replicate of qMS-normalized H3K36me2 and H3K27me3 CUT&RUN signal as indicated from D5 samples comparing distribution changes in response to NSD2i treatment. e, Representative tracks of qMS-normalized H3K36me2 and H3K27me3 CUT&RUN signal as indicated from D5 samples comparing distribution changes in response to NSD2i treatment at the indicated KRAS-effector genes. Chromosome location is indicated at the top and gene elements at the bottom, with exons as boxes and introns as lines.

Extended Data Fig. 5 |. NSD2i regulation of the epigenome.

a, Heatmap of H3K36me2 and H3K27me3 CUT&RUN signal across averaged gene body from data shown in Fig. 3g for each biological replicates. b, Average plot of H3K36me2 and H3K27me3 at intergenic regions. c, Average plot of indicated CUT&RUN plots across averaged gene body as in Fig. 3g.

Extended Data Fig. 6 |. NSD2i regulation of epigenomic and gene expression crosstalk.

a, GSEA of the indicated gene signatures of differential transcriptomes of MiaPaCa2, KP2 and NCI-H1373 cancer cell lines cells treated with 100 nM IACS17596 (see Supplementary Table 4). Circle sizes denote the normalized enrichment scores (NES), and the colour scale indicates the adjusted P value (P_adj) calculated by two-sided permutation test. b, Violin plot of gene expression of indicated genesets. Expression level shown in transcripts per million. Mean expression of each geneset is shown in red. P values determined by the two-sided Wilcoxon signed-rank paired test and based on the following n (number of genes for each functional group): PRC2 targets and cognate control gene number = 469; PRC2 neighbours and cognate control gene number = 738; PRC2 enhancer and cognate control gene number = 1000; KRAS signaling and cognate control gene number = 107; Pancreatic neoplasia and cognate control gene number = 304; EMT and cognate control gene number = 125 genes. c, Violin plot of MS-Normalized H3K36me2 CUT&RUN signal as in Fig. 4h. P values determined by the two-sided Wilcoxon signed-rank paired test with an n-number as in (b). d, Violin plot of NicE-seq density signal at gene promoters in the indicated gene set and conditions and n as in (b). P values determined by the two-sided Wilcoxon signed-rank paired test (b, c, d). Violin plots: boxes: 25^th to 75^th percentile, whiskers: min. to max., centre line: mean (b, c, d).

Extended Data Fig. 7 |. NSD2 inhibitors show acceptable drug-like ADME and pharmacokinetic properties and are well tolerated in vivo.

a, KRASi treatment does not impact levels of NSD2 or H3K36me2. Western analysis using the indicated antibodies of WCEs from the indicated cell lines treated for 1 day with 100 nM sotorasib. The asterisk denotes the shift in migration of KRAS^G12C covalently bound to sotorasib. H3 and tubulin are shown as loading controls. b, Bliss-Loewe synergy score calculations. Colony formation was assessed in a patient-derived LUAD cell line treated with varying concentrations of NSD2i (0 to 100 nM) and KRASi (sotorasib) (0–25 nM), which resulted in a Bliss-Loewe synergy score of 21.6 ( > 10 indicates synergy) (see Fig. 5a). c, Summary of in vitro ADME (absorption, distribution, metabolism, and excretion) properties of IACS-17596 and IACS-17817, see Methods. d, Summary of in vivo pharmacokinetic properties of IACS-17596 and IACS-17817 (left panel) and plasma concentration of NSD2i compounds following 100 mg/kg dose (oral gavage) quantified using liquid chromatography with tandem mass spectrometry (LC-MS/MS), see Methods. e, Western analysis with indicated antibodies of representative lysates of tissue biopsies from K^CPC PDAC mouse model treated with indicated concentrations of IACS-17817 for 8 days. Vinculin and total H3 shown as loading controls. Representative of n = 2 animals per treatment condition. f-g, Clinical pathology evaluations of wildtype mice treated with NSD2i (IACS-17817, 100 mg/kg, daily for 14 days, n = 6 mice) revealed no signs of significant toxicity: (f) body weight for each mouse was recorded after 14 days of treatment with NSD2i or control (vehicle) and the values are shown relative to weight measured at enrollment. (g) HE staining and immunohistochemistry (IHC) were performed on sections of the intestine, liver, kidney, and lung from animals as in (f) using the indicated antibodies. No significant differences in microscopic morphology or apoptotic and proliferative indices were observed between the treatment groups. Top panel: representative HE and IHC sections; scale bars indicate 100 μm. Bottom panel: quantification of proliferation (Ki67⁺) and cell death (cleaved Caspase3⁺) in the indicated conditions and represented in the top panel. P values determined by two-tailed unpaired t-test (f-g).

Extended Data Fig. 8 |. NSD2i treatment does not impact CBC and standard blood chemistry values or weight in mice.

a, Complete blood count and blood chemistry panels were determined for mice treated with NSD2i and vehicle control. P values determined by two-tailed unpaired t-test. b-e, Weight changes measured for each individual mouse over the course of the indicated treatments compared to weight at the enrollment for PDAC-PDX (b), LUAD-PDX (c), autochthonous PDAC (d) and LUAD (e) models (as in Fig. 5c–j).

Extended Data Fig. 9 |. Single-cell transcriptomics of K^C/CP tumors reveals oncogenic gene expression programming suppression and immune activation upon NSD2 and KRASi treatment.

a, Marker gene expression across defined cell types. Circle size is proportional to the percentage of cells expressing a gene and colour intensity is proportional to average scaled gene expression. b, UMAP of total single cell dataset split by treatment group. Shared nearest neighbour (SNN)-based clustering resolved transcriptomes into eight clusters, reflecting the different cell types. c, Projection of K^C/CP epithelial cells from each treatment group onto a LUAD cell atlas from KP-Tracer (left panel), shown separately as contour density plots. KP-tracer analysis indicates that K^C/CP control tumours are dominated by cancer cell populations exhibiting high-fitness transcriptional states characteristic of late-stage epithelial-mesenchymal transition-related clusters. In contrast, KRASi+NSD2i treated tumours were enriched for cells expressing low-fitness programs characteristic for early-stage alveolar type 2 (AT2)-like neoplastic transformation. d, Bhattacharyya distances showing transcriptomic variances between vehicle- and NSD2i-treated tumour or normal epithelial cells from scRNA-seq analysis of K^C/CP lung tumor biopsies. The box plot shows the median (centre line), interquartile range (IQR), and whiskers extending to 1.5× the IQR. P-value by two-sided Mann-Whitney-Wilcoxon unpaired test reported at calculation limit as <2.2 × 10⁻¹⁶. e. Heatmap depicting the expression levels of leading-edge genes enriched in the indicated GSEA signatures, comparing vehicle- and NSD2i-treated tumour cells and non-malignant normal epithelial cells from scRNA-seq analysis of K^C/CP lung tumour biopsies. Z-score normalized gene expression is shown.

Extended Data Fig. 10 |. Model of NSD2i mechanism of action for treatment of KRAS-driven and potentially other cancers.

a, Proposed model for how (i) the chromatin state at target pro-oncogenic genes, like those dependent on KRAS signaling, are regulated by NSD2 during tumorigenesis to influence transcriptional output and (ii) compounds that inhibit NSD2 and KRAS can synergize at the level of chromatin to suppress tumor growth. From left to right: (1) Normal tissue: transcription factors (TFs) that are activated by oncogenic signals (e.g. mutant KRAS) are normally dormant. Additionally, the chromatin at the TFs gene targets is in a low accessibility state due to H3K27me3 enrichment, likely facultative heterochromatin, which is established during tissue differentiation. Thus, transcriptional output of pro-oncogenic genes is minimal. (2) Tumour initiation: acquisition of GOF mutant KRAS (indicated here as KRAS*) locks mutant KRAS in the active state, resulting in constitutive activation of the previously dormant downstream pro-oncogenic TFs. These TFs induce transcription at pro-oncogenic gene targets (e.g. KRAS-dependent targets), but robust transcriptional activation is limited by the underlying repressive chromatin state. Thus, transcriptional output of pro-oncogenic genes is moderate. (3) NSD2-mediated malignant progression: in many cancers (e.g. see Extended Data Fig. 1b) NSD2 is upregulated. In such cancers, increased NSD2 activity deposits H3K36me2 throughout the genome, including at silenced, facultative heterochromatinized genic regions. The addition of H3K36me2 inhibits EZH2 and displaces H3K27me3, preventing the generation and spread of H3K27me3, increasing chromatin accessibility and therefore the ability of mutant KRAS-dependent TFs to robustly activate transcription of their target gene. Thus, transcriptional output of pro-oncogenic genes is high. In parallel, pathologic activation of the NSD2-H3K26me2 pathway, via chromatin opening, facilitates transcription of other oncogenic gene expression programs (e.g. EMT, Inflammation) that can cooperate with KRAS signalling to promote malignant progression. (4) NSD2i blocks H3K36me2 generation and thereby acts at the level of chromatin to revert the epigenomic landscape at oncogenic genes from a pathologic (open) to more physiologic (closed) chromatin state, reducing transcriptional output at pro-oncogenic genes. NSD2i, combined with KRASi, further reduces transcription of KRAS-dependent target genes, providing synergistic anti-tumour activity and durable tumour regression. The selectivity of NSD2i for pro-oncogenic genes lies in the chromatin at these genes having significantly higher H3K27me3 levels under baseline conditions; we speculate that in normal tissue these genes are present within facultative heterochromatin that formed during tissue differentiation and thus refer to them as H3K27me3-legacy loci. The higher baseline level of H3K27me3 renders the genes particularly sensitive to NSD2i-induced H3K27me3 elevation. b, Summary for how increased NSD2 activity, through increased expression or acquisition of GOF mutations, leads to aberrant epigenetic reprogramming that promotes tumorigenesis. Thus, the many cancers that have either NSD2 overexpression or harbour GOF mutations are predicted to be responsive to NSD2i treatment.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41586-025-09299-y.

Data availability

Plasmids, antibodies and cell lines generated in this study will be available from the lead contact upon request with a completed material transfer agreement. Custom scripts used to generate the results and figures are available from GitHub (https://github.com/andygglez/NSD2i_MIAPaCa). The following data are available from the Gene Expression Omnibus (GEO) database: raw and processed data of the scRNA-seq experiments (GSE273502); raw RNA-seq data (GSE273492); NicE-seq data (GSE273491) and CUT&RUN seq data (GSE273490). For human samples, the human genome assembly reference hg38 was used. For mouse samples, the mm10 genome assembly reference was used. NMR data have been deposited into the PDB with the identifier 9CVD. MS data have been deposited into the ProteomeXchange Consortium with the dataset identifier PXD059343. Source data are provided with this paper.