Dual inhibition of KDM4B and KDM5A disassembles the PAX3-FOXO1 transcriptional program in fusion-positive rhabdomyosarcoma

Abstract

Background

Fusion-positive rhabdomyosarcoma (FP-RMS) is driven by the oncogenic transcription factor PAX3-FOXO1 and is associated with poor clinical outcome. The histone demethylase KDM4B has been implicated in sustaining PAX3-FOXO1-dependent transcriptional programs, but the epigenetic regulatory mechanisms supporting this network remain incompletely characterized.

Methods

We employed a deep learning-based drug discovery strategy and identified a novel small-molecule, Compound 01. Functional and mechanistic analyses were then performed to delineate cooperative interactions among KDM4B, KDM5A, and PAX3-FOXO1, and to evaluate the impact of pharmacological perturbation on the stability of this transcriptional network in FP-RMS models.

Results

In this study, we identified Compound 01 with dual activity against KDM4B and KDM5A that triggers proteasome-dependent loss of both proteins. Mechanistically, we uncover a reciprocal stabilization loop in which KDM4B, KDM5A, and PAX3–FOXO1 reinforce each other to sustain the PAX3–FOXO1 transcriptional state. Pharmacological disruption of this axis by Compound 01 collapses the PAX3–FOXO1-driven oncogenic transcriptional program, suppresses key downstream targets, and markedly impairs FP-RMS cell proliferation and metastatic potential.

Conclusion

These findings identify KDM5A as a critical epigenetic partner of the PAX3–FOXO1 network and establish dual KDM4B/KDM5A targeting as a strategy to destabilize PAX3–FOXO1 transcriptional circuitry in FP-RMS.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13062-026-00837-w.

Introduction

Rhabdomyosarcoma (RMS) is the most common soft tissue sarcoma in children [1, 2]. It is classified into two major histological subtypes: alveolar RMS (aRMS) and embryonal RMS (eRMS). The origin cells of aRMS are thought to derive from fetal myogenic cells and are driven to malignant transformation by specific fusion oncogenes, PAX3-FOXO1 or its variant PAX7-FOXO1 [3–5]. PAX3-FOXO1 fusion-positive RMS (FP-RMS) is more aggressive, with higher metastatic rates and poorer prognosis compared to fusion-negative RMS (FN-RMS) [1, 6, 7]. Although current therapies have steadily improved survival rates for low-risk RMS patients, outcomes for high-risk patients, especially those with fusion-positive RMS, remain unsatisfactory [8]. As an oncogenic transcription factor, PAX3-FOXO1 plays a critical role in the initiation, proliferation, and survival of aRMS. It regulates the expression of various oncogenes by inducing reprogramming of the cis-regulatory landscape via super-enhancer formation. Recent studies have further defined the role of PAX3-FOXO1 as a transcription factor, demonstrating its crucial function within a transcriptional network that guides myogenic programs, alongside MYOD1, MYOG, MYCN, and SOX8 [9–11]. However, directly targeting PAX3-FOXO1 remains a formidable challenge due to its lack of enzymatic activity and the absence of defined small-molecule binding pockets, characteristics that have historically rendered it “undruggable”. Consequently, despite preclinical models and clinical trials testing therapies targeting PAX3-FOXO1 downstream molecules [12–16], effective and safe strategies for targeting the transcriptional network controlled by PAX3-FOXO1 have yet to be established.

Recent studies have shown that histone lysine demethylases (KDMs), key enzymes in epigenetic regulation, are frequently dysregulated in cancers driven by fusion (chimeric) transcription factors [17–21]. In rhabdomyosarcoma (RMS), for example, the PAX3-FOXO1 controls a core transcriptional circuitry, and the histone demethylase KDM4B is an essential component of this network: genetic or pharmacological inhibition of KDM4B markedly impairs PAX3-FOXO1 function, downregulating its downstream target genes and eliciting potent anticancer effects [22]. These results indicate that KDM4B sustains the PAX3-FOXO1-driven transcriptional circuit and that targeting KDM4B is a promising therapeutic strategy for fusion-positive RMS. Likewise, KDM3B has been implicated in PAX3-FOXO1-mediated oncogenesis: its suppression robustly downregulates PAX3-FOXO1 target genes and activates a skeletal muscle myogenic differentiation program with accompanying apoptosis [23]. Together, these findings suggest that the family of histone demethylases plays an active pro-tumorigenic role in RMS. Intriguingly, FP-RMS tumors harbor essentially no recurrent point mutations [24, 25], yet FP-RMS has a far worse prognosis than FN-RMS [6]. This paradox underscores the critical importance of epigenetic mechanisms in FP-RMS pathogenesis.

Ligand-based virtual screening (LBVS) is a widely applied computational drug discovery approach [26]. In this study, we used two in-house deep learning-based molecular representation models, GeminiMol and PhenoModel, which incorporate molecular conformational space into their embeddings, enabling more comprehensive and fine-grained molecular feature characterization and supporting ligand similarity-based drug discovery [27, 28]. Such representations are particularly advantageous for downstream tasks such as LBVS, as they leverage the chemical and biological properties of known active ligands to identify potential compounds with high likelihood of interacting with the biological target. The models have been demonstrated to enhance ligand-based drug discovery by improving the ability to capture conformational diversity and thereby increasing both predictive performance and drug-likeness of candidate molecules.

In this study, we collected five molecules with known inhibitory activities against KDM4B (Supplementary Table 1). Using GeminiMol and PhenoModel, we encoded these molecules and constructed a pharmacophore profile by weighting molecular features according to their activity. Three commercial compound libraries containing a total of approximately 18 million compounds were subsequently screened, and candidate molecules were ranked based on the similarity scores between the encoded molecular representations and the constructed pharmacophore profile. The top-ranked compounds were further validated using receptor-based virtual screening. Through this integrative approach, we identified a new class of KDM4 inhibitors that bind to the enzymatic pocket of KDM4B, thereby suppressing its demethylase activity. This inhibition led to reduced expression of PAX3-FOXO1 at both the transcriptional and protein levels, resulting in the downregulation of its downstream targets and exerting potent tumor-suppressive effects. Furthermore, we discovered that KDM5A also serves as a target of these compounds, functioning in concert with KDM4B and PAX3-FOXO1 to form an interaction network that drives oncogenic pathways. Collectively, these findings reveal a cooperative mechanism whereby KDM4B and KDM5A promote the transcriptional regulation of PAX3-FOXO1 target genes to sustain FP-RMS growth, and highlight a new molecular scaffold for inhibiting both KDM4 and KDM5 family proteins.

Methods

Identification of inhibitors against KDM4B

Ligand-based drug discovery approaches leverage multi-view molecular similarities, including pharmacophoric and conformational characteristics, to identify compounds with comparable bioactivities yet distinct chemical scaffolds, thereby offering a powerful strategy for scaffold hopping. To identify compounds similar to the template, we used our in-house deep learning-based molecular representation model GeminiMol, which encodes bioactive conformational space, together with PhenoModel, which integrates cellular phenotype statistics with molecular structures. This strategy enables ligand-based virtual screening to identify potential hits with similar structural, conformational, and phenotypic characteristics, improving the efficiency of novel compound discovery. The template compounds used for screening comprised a series of previously reported KDM4B enzymatic inhibitors with relatively high demethylation efficacy, collected from the literature (compound structures are shown in Fig. 1A). Candidate compounds identified based on structural similarity were further evaluated by molecular docking to assess their binding to the target protein.

Fig. 1.

Discovery and validation of KDM4B active compounds and the inhibition effects against PAX3-FOXO1. (A) Compound Screening Workflow: Active molecules targeting key binding sites of KDM4B are collected from literature as reference compounds for the input model. Utilizing GeminiMol and PhenoModel, these are applied to a LigPrep-processed compound library to identify potential inhibitors. Subsequently, the validation of binding mode is conducted through the implementation of molecular docking and SPR experiments, with the objective of refining and prioritizing the candidate compounds. CCK-8 is further employed for the purpose of cellular-level activity testing of candidate compounds. Furthermore, the laboratory’s non-virtual compound library undergoes preliminary screening via SPR and is subsequently incorporated into the candidate compound list for consideration. (B) Chemical structures of the five active compounds. (C) Dose-response curves of cell viability in RH30, RH41 and RD cell lines after 72 h treatment of the compounds and SPR sensorgrams tested by CCK-8 assay and Biacore 8 K, respectively. (D) Western validation of fusion protein PAX3-FOXO1 in RH30 cell lines after 72 h treatment with Compound 01. (E) Real-time quantitative PCR of PAX3-FOXO1 in RH30 cell lines after 72 h treatment with Compound 01

Compounds from three commercial databases, Specs (https://www.specs.net/), MedChem Express (MCE) (https://www.medchemexpress.cn/) and TargetMol (https://www.targetmol.com/) were used and prepared using the LigPrep tool by Schrödinger. To facilitate the screening process, we constructed a weighted pharmacophore model based on the bioactivity of the reference molecules. Specifically, weights were assigned to the pharmacophore features of each template molecule according to its potency, where features derived from highly active compounds were given higher weights to prioritize their significance in the molecular search. This approach enables the identification of candidate molecules with higher-weighted features, thereby improving the success rate of the screening. Molecules with high activity and selectivity were given higher weights, while molecules with poor selectivity were assigned negative or penalizing weights. Subsequently, a high-similarity enrichment process was conducted to identify potential active molecules from three commercial compound databases. The top 1,000 molecules ranked by each model (GeminiMol and PhenoModel) were individually docked into the binding pocket of KDM4B, and their binding modes were assessed using Glide. Based on the integrated evaluation of deep learning scores, docking scores, and visual inspection for optimal binding interactions, 20 candidate compounds were finally selected for further in vitro experimental validation.

Molecular docking

The compounds were docked into KDMs’ x-ray structures using Schrödinger Release 2019-1 (Protein Preparation Wizard, Epik, Impact, Prime, Schrödinger, LLC, New York, NY, 2019). The following structures from the Protein Data Bank (PDB) have been used in docking: PDB: 7JM5 for KDM4B; PDB: 3GL6 for KDM5A. The other structures used in the study were supplemented in Supplementary Materials. The filtered small molecules from initial screening were prepared using LigPrep [29]. Glide was employed for structure-based molecular docking to analyze binding patterns [30]; the docking was conducted at standard precision (SP) without enhanced sampling.

In-house compound library screening

The laboratory’s in-house compound library includes over 1,000 compounds. These compounds are stored at a concentration of 5 mM in DMSO and are stored in 384-well plates. We employed Surface Plasmon Resonance (SPR) for high-throughput affinity screening of this compound series targeting recombinant KDM4B. The compounds were aliquoted into 384-well plates and initially screened at a concentration of 50 µM. The SPR experiment parameters included a contact time of 60 s and a dissociation time of 60 s. The top 100 compounds ranked by SPR response values were further validated for their binding affinity using a concentration gradient in SPR assays.

Surface plasmon resonance (SPR) analysis

Before determining the binding affinity of each compound, recombinant KDM4B protein was diluted to 50 µg/mL in sodium acetate buffer (pH 4.5) and immobilized onto a Series S CM5 sensor chip (Cytiva, Marlborough, MA, USA) via standard amine coupling to achieve a level of 13,000 response units (RU). The compound stock solutions (100 mM) were diluted in 1× HBS-EP buffer (20 mM HEPES, pH 8.0, 150 mM NaCl, 3.0 mM EDTA, and 0.1% [v/v] Tween-20) to create a nine-concentration series ranging from 0.098 µM to 50 µM, with a final DMSO concentration of 5%. Three startup cycles were performed using running buffer. Subsequently, analytes were injected and passed through the chip, with an association time of 120 s and a dissociation time of 100 s. After each cycle, the chip was washed with 50% DMSO. Solvent correction was performed every 48 cycles, accounting for DMSO concentration variations between 4.5% and 5.8%. The raw data were processed, double-referenced, and solvent-corrected using Biacore 8 K Evaluation Software, and the KD values for each compound were determined using a steady-state affinity model with a constant Rmax.

Cell culture

Human HEK293T were obtained from Cell Bank, Chinese Academy of Sciences (Shanghai, China). Human RH30 and RD rhabdomyosarcoma cell lines were purchased from Sunncell (Wuhan, China). RH41 rhabdomyosarcoma cell line was a gift from Jinhu Wang (Zhejiang University). HEK293T, RH30 and RH41 cells were cultured in RPMI 1640 (MA0215, MeilunBio, Dalian, China). RD cells were cultured in DMEM (MA0213, MeilunBio, Dalian, China). The base medium was supplemented with 10% (v/v) FBS (Gibco), 100 U/mL penicillin and 100 µg/mL streptomycin (Gibco). Cells were maintained at 37 °C in an atmosphere of 5% CO₂.

Cell proliferation assay

Cell proliferation was measured using the Cell Counting Kit-8 (CCK-8) (AC11L054, Life-iLab, China) according to the manufacturer’s protocol. Cells were seeded into 96-well plates at a density of 6 × 10³ (RH30, RH41), 5 × 10³ (RD) cells per well. After 24 h incubation, the medium was replaced with fresh medium containing either 0.1% DMSO (control) or different concentrations of the compound (0.78, 1.56, 3.125, 6.25, 12.5, and 25 µM), with a final DMSO concentration of 0.1%. After 72 h of incubation, cell viability was assessed. CCK-8 reagent (10 µL) was added to each well and incubated at 37 °C for 1 h. A microplate reader (PerkinElmer, Waltham, MA, USA) was used to measure the absorbance values at 450 nm. GraphPad Prism software (GraphPad Software, 10.1.2, San Diego, CA, USA) was used to generate EC₅₀ values. Error bars on proliferation curves represent the standard errors of the mean.

Total RNA extraction, cDNA synthesis, and real-time PCR

Total RNA extraction from cells was performed using the AFTSpin Tissue/Cell Fast RNA Extraction Kit (RK30120, ABclonal, Wuhan, China) according to the manufacturer’s instructions. cDNA was synthesized from total RNA using ABScript II cDNA First-Strand Synthesis Kit (RK20400, ABclonal, Wuhan, China) according to the manufacturer’s instructions. Quantitative real-time polymerase chain reaction (qRT-PCR) was performed using ABScript II One Step SYBR Green RT-qPCR Kit (RK20404, ABclonal, Wuhan, China). The results were analyzed using the ΔΔCT methods. The PCR primer sequences are listed in the Supplementary Materials.

SDS-PAGE and western blot

Cells were washed twice with ice-cold phosphate-buffered saline (PBS) and directly lysed on ice with 1X sample loading buffer (0.1 M Tris HCl [pH 6.8], 200 mM dithiothreitol [DTT], 0.01% bromophenol blue, 4% sodium dodecyl sulfate [SDS] and 20% glycerol). On ice, cell lysates were collected and followed by 10 min heating at 95 °C. After the cell lysates were briefly centrifuged at 13,000×g at room temperature for 5 min, 10–20 µl of the cell lysates were separated on FuturePAGETM 4–20% 15 Wells (ET15420LGel, ACE) and transferred to methanol-soaked polyvinylidene difluoride (PVDF) membranes (Millipore). Membranes were blocked in TBS buffer supplemented with 0.1% TWEEN 20 (TBS-T) and 5% skim milk and incubated for 1 h at room temperature under gentle horizontal shaking. Membranes were incubated overnight at 4 °C with the primary antibodies under gentle horizontal shaking. The primary antibodies were prepared in TBS-T or Universal Antibody Dilution Buffer (Epizyme, Shanghai, China) with the antibodies and respective dilutions listed in the Supplementary Materials.

The next day, membranes were washed 3 times (for 5 min) with TBS-T at room temperature under gentle horizontal shaking. While shielded from light, membranes were then incubated with goat anti-mouse (H + L) or goat anti-rabbit HRP-conjugated secondary antibodies (H + L) for 1 h at room temperature under gentle horizontal shaking, followed by three 5-minute washes with TBS-T at room temperature. Proteins were visualized with Enhanced Chemiluminescence Kit (EpiZyme, Shanghai, China) and detected by Amersham Imager 680 (GE) and e-BLOT touch imager (Shanghai, China).

Cellular thermal shift assay (CETSA)

RH30 cells were seeded in 10 cm dishes and treated with either DMSO or 10 µM Compound 01 for 24 h. After treatment, cells were harvested, washed three times with ice-cold PBS, and resuspended in PBS supplemented with a protease inhibitor cocktail. The cell suspensions were then divided into seven equal aliquots and transferred into PCR tubes. These aliquots were subjected to a thermal challenge at specific temperatures (37, 42, 47, 52, 57, 62 and 67℃) for 5 min using a thermal cycler, followed by a 3 min incubation at room temperature. To obtain the soluble protein fraction, the cells were lysed by five consecutive cycles of freeze-thaw using liquid nitrogen and a 37℃ water bath. The lysates were then centrifuged at 20,000 × g for 30 min at 4℃ to precipitate the denatured and aggregated proteins. The resulting supernatants, containing the stabilized soluble proteins, were collected and analyzed by Western Blotting using an anti-KDM5A antibody.

Protein expression and purification

The pET28a expression vector encoding KDM4B (residues 1-366) with N-terminal 6× His tag was transformed into Rosetta™ 2(DE3) Competent Cells (TransGen Biotech, Beijing, China). Cells were grown in LB medium at 37 °C until OD600 = 1.0 and protein was expressed at 16 °C for 20 h after 0.2 mM isopropyl-β-D-thiogalactoside (IPTG) was added. After collecting by centrifugation, cells were suspended with lysis buffer (50 mM Tris, pH 8.0, 0.5 M NaCl, 10 mM imidazole, 1 × EDTA-free protease inhibitor cocktails (C0001, TargetMol). After high pressure crushing, the lysate was cleared by centrifugation at 4 °C, and the supernatant was collected and incubated with Ni2+-NTA beads (Smart-Lifesciences, Changzhou, China) for 60 min. The mixture was loaded into column and washed with wash buffer containing 60 mM imidazole and eluted with buffer containing 300 mM imidazole. Fractions were further purified using size exclusion chromatography (Superdex 200 Increase 10/300, GE, Fairfield, CO, USA) with buffer of 50 mM HEPES, pH 8.0, 0.5 M NaCl. Protein was concentrated to 2 mg/mL for further investigation.

RNA-seq

RH30 and RD cells were seeded into 10 cm plates at a density of 4 × 10⁶ cells per plate. After a treatment of 10 µM compound 01 or 0.1% DMSO only for 72 h, cells were washed in PBS then lysed in 500 µL Trizol. RNA isolation, library preparation and RNA-seq was performed by LC-Bio Technology (Hangzhou, China) using NovaSeq 6000 platform. Prior to sequencing, RNA integrity was assessed by 2100 Bioanalyzer and agarose gel electrophoresis.

Gene Set Enrichment Analysis (GSEA) was performed on log2-fold change ranked list calculated by comparison with control in samples with single experiments using Run GSEAPreRanked with default settings [31].

Cell cycle analysis

Cell cycle was monitored using BD Cycletest™ Plus DNA Reagent Kit (BD Biosciences, Franklin Lakes, NJ, USA) in accordance with the manufacturer’s protocol and flow cytometry. RH30 and RD cells were seeded into 10 cm plates at a density of 4 × 10⁶ cells per well. Cells were trypsinized and harvested after being treated with 10 µM compound 01 for 72 h. Followed by washing with PBS, solution A, B, and C of the kit was added to stain the cells with PI. The PI signal was measured by flow cytometry. The results were analyzed by FlowJo software.

Apoptosis assay

Cell apoptosis was quantified using FITC Annexin V Apoptosis Detection Kit (BD Biosciences) following the manufacturer’s protocol and flow cytometry. Cells were collected after a 72 h treatment of 10 µM Compound 01 and washed twice with cold PBS. Subsequently, cells were resuspended in 1× Binding Buffer and mixed with FITC Annexin V and PI. Cells were subjected to a flow cytometer (Beckman Coulter, Brea, CA, USA) after an incubation for 15 min in the dark. The results were analyzed by ModFit software.

Immunofluorescence staining

For cell immunofluorescence staining, cells on 20-mm diameter glass coverslips (NEST) were fixed with 4% paraformaldehyde at room temperature for 10 min and permeabilized with 0.3% Triton X-100 for 10 min. Slides or coverslips were blocked with 3% BSA for 1 h and then incubated with the primary antibodies overnight at 4 °C. After washing, slides or coverslips were incubated with secondary antibodies at room temperature for 1 h and then incubated with DAPI (RM02978, ABclonal) for 10 min. Finally, slides or coverslips were washed with PBS, mounted and then imaged with confocal microscope (Zeiss LSM 710).

Plasmids

For KDM4B and KDM5A overexpression experiments, RH30 cells were transduced with Flag-KDM4B (1-366aa)-Puro-Plvx and Flag-KDM5A (1-797aa)-Puro-Plvx, which was generated by Genscript (Nanjing, China). For KDM4B and KDM5A rescue experiments in CRISPR knockout cells, Rh30-KO-KDM4B and Rh30-KO-KDM5A cells were respectively transduced with Flag-KDM4B (1-366aa)-Puro-Plvx and Flag-KDM5A (1-797aa)-Puro-Plvx, which were generated by Genscript (Nanjing, China).

Generation of KDM4B, KDM5A CRISPR knockout cell lines

The sgRNA oligos used are listed in the Supplementary Materials.

sgRNA lentivirus assembly and transfection

293T cells were seeded in 10 cm dishes at 70–80% confluence the day before transfection. On the following day, when the cell density reached 90%, transfection was performed. The following preparations were made:

Tube 1: Prepare a mixture of KDM4B/KDM5A sgRNA: pspAX2 plasmid: PMD2G plasmid in a mass ratio of 4:3:1, with a total mass of 24 µg for three plasmids, and complete to 500 µL with opti-MEM.

Tube 2: Prepare the PEI transfection reagent, with PEI (1 mg/mL) at a 1:2 ratio (48 µg PEI) and complete to 500 µL with opti-MEM (11058021, Thermo/Life/invitrogen).

Slowly add the solution from Tube 2 to Tube 1 and incubate at room temperature for 20 min. After incubation, slowly add the suspension dropwise to the 293T cells in a clockwise motion. Continue culturing for 48 h. After 48 h, collect 8 mL of the 293T cell culture medium to harvest the transfected virus. Add 8 mL of fresh medium and continue culturing the 293T cells for another 24 h. After 24 h, collect 8 mL of the culture medium again to harvest the transfected virus. The virus was then concentrated using EZ Lentivirus Concentration Solution (AC04L442, Life-iLab) following the steps: Collect the 293T cell culture supernatant at 48 h and 72 h post-transfection. Centrifuge at 4 °C, 2000 g for 10 min; Filter the supernatant through a 0.45 μm low-protein-binding membrane to remove cell debris; Mix the viral filtrate with the concentration reagent at a ratio of 4:1 (v/v), shake well, and incubate at 4 °C overnight; After incubation, centrifuge at 4 °C, 3500 g for 15 min and carefully discard the supernatant; Gently resuspend the viral pellet in an appropriate volume of PBS by pipetting up and down, and use the concentrated virus to infect target cells.

Viral infection of RH30 cells

Screening of Puro Concentration in Target Cells: Seed the target cells in 12-well plates at a density of 80–90%. After cells adhere, increase the puromycin concentration (1–10 µg/mL per well) to identify the concentration that results in complete cell death, which will be used for subsequent selection; Seed target cells in 24-well plates at a density of 70–80% the day before transfection; After 24 h, when the cells have adhered, add 500 µL of the KDM4B/KDM5A viral concentrate and 0.8 µL (8 µg/mL) polybrene (40804ES86, Yeasen), and continue incubation for 12–18 h; After 12–18 h, replace the medium with fresh culture medium and continue incubation until the cells are confluent. Then, transfer to a 6-well plate for continued culture; After 24 h, when cells have stabilized and grown, add the puromycin concentration determined earlier to select for successfully infected cells with KDM4B/KDM5A knockdown; Once the cells are confluent, transfer them to a 10 cm dish for further culture. Perform Western blotting to verify successful knockdown.

Immunoprecipitation

Cells were washed twice with ice-cold phosphate buffered saline (PBS) and then directly lysed on ice with Cell lysis buffer for IP (RM00022, ABclonal), supplemented with 1×EDTA-free protease inhibitor cocktails (C0001, TargetMol). The cell lysate was transferred to a 1.5 mL centrifuge tube and incubate at 4 °C for 20 min, gently vortexing at 15–20 rpm. After centrifugation at 13,000 × g at 4 °C for 10 min, the pre-cleared supernatant was incubated with rotation at 4 °C overnight with 3 µg of anti-KDM4B (A6670, ABclonal), anti-KDM5A (# ET7107-43, HUABIO) and 3 µg normal rabbit anti-IgG antibody (AC005, ABclonal) as a negative control. The next day, 20 µl of protein A/G magnetic beads (RM02915, ABclonal) were washed 3 times at room temperature with the IP buffer and then added to each pre-cleared supernatant for 2-hour rotation at 4 °C. The supernatant (flow-through) was discarded, and the beads were washed 3 times with IP buffer, eluted with 35 µl of the 2X sample loading buffer (0.1 M Tris HCl [pH 6.8], 200 mM dithiothreitol [DTT], 0.01% bromophenol blue, 4% sodium dodecyl sulfate [SDS] and 20% glycerol) and heated for 10 min at 95 °C. Input lysate was heated for 10 min at 95 °C and 10 µl of co immunoprecipitation lysate and 10 µl of input reactions were run on FuturePAGETM 4–20% 15 Wells (ET15420LGel, ACE) and immunoblotted with anti-KDM4B (A6670, ABclonal), anti-KDM5A (# ET7107-43, HUABIO) and anti-FOXO1 (A25295, ABclonal) antibodies. Further Mass Spectrometry and proteomics were performed at the Majorbio (Shanghai, China).

siRNA transfection

siRNA transfection was performed using Lipofectaminerm 3000 Transfection Reagent (L3000015, Thermo Fischer Scientific) and NanoTrans Transfection Reagent 3000 (CT0006, Cytoch) according to the manufacturer’s instructions. The siRNA target sequences were purchased from Genscript (Nanjing, China), and the sequences are listed in the Supplementary Materials. After 72–96 h, cells were harvested for Western blot analysis.

Crystal violet staining

After removing media, cells were washed with phosphate buffered saline without calcium or magnesium (PBS) and treated with 4% formaldehyde for 20 min. Once 4% formaldehyde was removed, cells were stained with 0.1% crystal violet stain for 1 h.

Results

Initial screening based on LBVS

Given the large number of compounds in the screening library, we employed a deep learning-driven high-throughput virtual screening method for efficient initial screening. Based on several known efficient KDM4B inhibitors with different structures (Supplementary Table 1) as reference molecules, we performed an initial screening of the compound library using GeminiMol [27] and PhenoModel [28]. The affinity information of the reference molecules was used to assign scores when inputting them into the model. The screened molecules were ranked based on the similarity scores from both GeminiMol and PhenoModel, with the top 1,000 molecules from each model proceeding to the next stage of evaluation (Fig. 1A).

Refining the selection of candidate compounds based on molecular docking

The second step involved refining the selection of candidate compounds through molecular docking, where the final candidates were selected based on their interaction binding modes. Using the crystal structure from RCSB PDB (PDB ID: 7JM5) as a template for the identified binding site, we performed structure-based molecular docking to analyze the potential interaction modes between the compounds and KDM4B protein. The docking was carried out using the Glide module from Schrödinger to evaluate the binding modes [32]. The selection process for candidate compounds focused on the following criteria: compound-pocket adaptability, hydrogen bonding, hydrophobic interactions, chelation of the metal ion within the binding pocket, and relative drug-likeness. Based on this, we selected the top several dozen compounds with the best combination of GeminiMol and PhenoModel similarity scores and pocket interaction evaluations for further SPR affinity and anti-tumor cell activity validation (Fig. 1A) (Supplementary Table 2).

Assessment of compound-KDM4B binding capability and cell proliferation ability in RMS

We further evaluated the binding affinity of the compounds with KDM4B and their anti-tumor cell activity through SPR and CCK-8 assays. The existing in-house compound library was also included in the SPR affinity measurements. Among the compounds in the library, 16 compounds showed relatively strong binding to KDM4B, with K_D values ranging from 9.42 × 10^− 7 M to 8.75 × 10^− 5 M (Figures S1-2). These compounds were further tested for their anti-tumor cell activity.

CCK-8 assays revealed that 6 compounds inhibited the proliferation of FP-RMS cells while 10 compounds inhibited the proliferation of FN-RMS cells in a dose-dependent manner (Figures S3A-B). Given the aggressive clinical nature and poor prognosis of FP-RMS, we prioritized candidates based on their inhibitory effect within the FP-RMS subtype. Specifically, based on the binding mode predicted by molecular docking, compounds 01 (AK-778/41314756), 02 (AK-778/43206351), 03 (AK-778/43464903), 04 (AK-778/43420905), and 05 (AN-652/13035575) were selected as the top candidates with strong affinity for KDM4B and potent anti-cell proliferation activity (Figs. 1B-C).

Compound 01 inhibits the expression of PAX3-FOXO1

KDM4B has been shown to be closely associated with the stability of the fusion protein PAX3-FOXO1 [22]. After treatment with the candidate compounds, Compound 01 was found to decrease the fusion protein expression at a lower concentration (Fig. 1D, Figure S3C). We measured the transcriptional levels of PAX3-FOXO1 using q-PCR. The compounds were found to reduce the transcriptional levels of the fusion protein, with Compound 01 showing the most significant decrease (Fig. 1E, Figure S3D). Given the stronger anti-tumor cell proliferation ability and binding affinity to KDM4B, as well as the lower cytotoxicity (Figure S4A) and its stronger ability to reduce fusion protein expression in both RH30 and RH41 fusion positive cell lines (Figure S4B), Compound 01 was selected as a representative compound for further studies to reduce off-target risks and enhance targeting specificity to the fusion protein.

Compound 01 inhibits the demethylation function of KDM4B and KDM5A, which are common targets of the compound

To investigate the anti-tumor mechanism of Compound 01, we first assessed its inhibition of KDM4B demethylase function. H3K9Me3/2 and H3K36Me3/2 are the major epigenetic targets for demethylation by KDM4 family proteins [33]. In wild-type RH30 cells, treatment with Compound 01 alone led to a mild increase in the H3K9me3 levels (Figure S4C). After overexpressing KDM4B in RH30 cells to reduce the baseline levels of H3K9Me3, Compound 01 enhanced the levels of H3K9Me3 obviously (Fig. 2A), suggesting effective inhibition of KDM4B demethylase function. An increase in H3K36Me2 levels was also observed (Figure S4C). Through molecular docking, we showed that Compound 01 can bind to the demethylase pocket of KDM4B, forming a coordination bond with the metal iron ion inside the pocket (Fig. 2B), stabilizing the compound’s binding mode, which is consistent with the results of its inhibition of demethylase activity.

Fig. 2.

KDM4B and KDM5A serve as the common targets of Compound 01. (A) Western blot of RH30 cells treated with 10 µM Compound 01 for 72 h after transfected with KDM4B overexpression plasmid showing increases in methylation of H3K9me3. (B) Molecular docking results demonstrate the binding mode of Compound 01 within the active site pocket of KDM4B and KDM5A. (KDM4B: blue; KDM5A: red) (C) Western blotting validated the efficiency of KDM4B knockout by CRISPR-Cas9 in RH30 cells. The effect of Compound 01 was attenuated in RH30 cell lines after KDM4B knockout. (D) Western blotting validated the efficiency of KDM5A knockout by CRISPR-Cas9 in RH30 cells. The effect of Compound 01 was attenuated in RH30 cell lines after KDM5A knockout. (E) Western blotting validated the efficiency of KDM6B knockout by CRISPR-Cas9 in RH30 cells. The effect of Compound 01 wasn’t attenuated in RH30 cell lines after KDM6B knockout. (F) Western validation of the histone substrates for KDM5A and KDM6B demethylation in RH30 cell lines after 72 h treatment with Compound 01. (G) Western blotting demonstrated the degradation of KDM4B and KDM5A after 72 h treatment with Compound 01 in indicated concentration. (H) Real-time quantitative PCR of PAX3-FOXO1 in RH30 cell lines after 72 h treatment with Compound 01 in 15 µM showing unchanged mRNA expression level of KDM4B and KDM5A. (I) Western-blot showing that KDM5A downregulation by 10 µM of Compound 01 treatment was rescued by addition of 10 µM of MG132 to the culture medium in RH30 cells. (J) Western-blot showing that KDM4B downregulation by 10 µM of Compound 4756 treatment was not rescued by addition of 10 µM of MG132 to the culture medium in RH30 cells

We then established a KDM4B knockout cell line to further evaluate the inhibitory efficacy of compound. The knockout of KDM4B weakened the effect of Compound 01, validating that KDM4B is an essential target of Compound 01 (Fig. 2C).

To verify and further explore the mechanism of Compound 01, we included KDM5A in our validation analyses, given its higher structural similarity to the KDM4 demethylase pocket and its relatively high expression in RMS (Fig. 2B, Figure S4D) [34]. In addition, given that KDM6B serves as a major regulator of H3K27me3 and exhibits comparatively high expression within the KDM6 family in RMS [35] (Figure S4D), we also incorporated KDM6B into our validation experiments.

We examine its inhibitory effect on the demethylase activity of KDM5A and KDM6B by measuring the levels of H3K4Me3/2 and H3K27Me3, the substrate for KDM5 and KDM6 family demethylases [36]. Compound 01 increased H3K4Me3 levels, indicating effective inhibition of KDM5A demethylase activity (Fig. 2F). However, there was no increase in H3K27Me3, a substrate for KDM6 family demethylases was observed (Fig. 2F), and other histone marks also remained unaffected (Figure S4C). We also established KDM5A and KDM6B knockout cell lines, genetic ablation of KDM5A attenuated the effect of Compound 01, confirming that KDM5A is a critical molecular target of Compound 01 (Fig. 2D). However, the knockout of KDM6B did not weaken the effect of Compound 01, indicating that KDM6B is not a target of the compound (Figs. 2E). To test if KDM5A is directly engaged by Compound 01 within cellular environment, we performed a Cellular Thermal Shift Assay. Our results demonstrated that Compound 01 treatment increased the thermal stability of KDM5A in RH30 cell lysates (Figure S4E), providing evidence of target engagement.

Compounds like QC6352 have been shown to directly reduce the protein levels of the target KDM4B through the proteasomal pathway [22]. In this study, we aimed to determine whether Compound 01 also reduces the protein levels of its target. Western blot analysis revealed that Compound 01 significantly reduced the expression of KDM5A and KDM4B in fusion-positive RH30 and RH41 cells (Fig. 2G, Figure S4F), and the rest of candidate compounds also showed respectively down-regulation (Figure S4G). However, in the fusion-negative RD cell line, the reduction of KDM5A and KDM4B also exist but is less pronounced compared to that observed in the fusion-positive cell line. (Figure S4H). We also assessed the expression of other KDM family proteins and found that their levels were unaffected by Compound 01 (Figure S4I-J), suggesting that the compound’s targeting is selective. Molecular docking showed that among members of the KDM family, KDM4B and KDM5A had the highest docking scores with Compound 01, consistent with our experimental results (Figure S4K).

Furthermore, no changes in the transcriptional levels of KDM5A and KDM4B were observed (Fig. 2H), indicating that the compound may induce protein degradation through the proteasomal pathway. To confirm this, we added the proteasome inhibitor MG132 to the culture medium, which rescued the expression of KDM5A proteins (Fig. 2I). However, we did not observe significant re-regulation of the KDM4B protein, suggesting that Compound 01 may have affected the protein stability of KDM4B in different way than KDM5A (Fig. 2J).

KDM4B and KDM5A drive cell proliferation, migration and progression in RMS, and suppressed by compound 01

Research on the KDM5A family in RMS and other cancers is still limited. However, considering the important role of the KDM family in the epigenetic regulation of several cancers in recent years [37–41], we hypothesize that KDM5A also plays a critical role in RMS. Studies conducted on KDM knockout cell lines revealed that both KDM4B and KDM5A knockout inhibited the proliferation and reduced colony formation, and Compound 01 replicates the inhibitory phenotype (Figs. 3A-B, Figure S5A-B). Reintroducing KDM4B and KDM5A expressions (Fig. 3C) largely rescued cell proliferation, confirming that the observed phenotype is specific to KDM4B and KDM5A (Fig. 3D).

Fig. 3.

KDM4B and KDM5A are required for cell proliferation and tumorigenesis of FP-RMS Cells. (A) Growth curve of parental RH30 cells, KDM4B and KDM5A knockout RH30 cells and RH30 cells treated with Compound 01 in indicated concentration gradient. (B) KDM4B, KDM5A knockout clones and parental RH30 cells were seeded at a low cell density (1000 cells per well) in triplicates. The cells were cultured for 14 days and stained by crystal violet. (C) Western validation subsequent with respectively introduction of KDM4B and KDM5A overexpression plasmid vectors into the KDM4B, KDM5A knockout clones. (D) KDM4B, KDM5A knockout RH30 clones, and KDM4B, KDM5A reintroduction clones were seeded at a low cell density (1000 cells per well) in triplicates. The cells were cultured for 10 days and stained with crystal violet. (E) Representative images showing the migration of KDM4B, KDM5A knockout RH30 cells and RH30 parental cells across the scratch wound at 0, 24, 48 h post a linear scratch made by a sterile pipette tip. (F) Quantitative bar chart and line chart of the migration area of KDM4B, KDM5A knockout clones versus parental RH30 cells. Statistics calculated using biological triplicates. ***, p < 0.0005, ****, p < 0.00005. (G) Western validation of the respectively introduction of KDM4B and KDM5A overexpression plasmid vectors into the RH30 cells. (H) Representative images showing the migration of KDM4B, KDM5A overexpression RH30 cells and RH30 parental cells across the scratch wound at 0, 24, 48 h post a linear scratch made by a sterile pipette tip. (I) Quantitative bar chart and line chart of the migration area of KDM4B and KDM5A overexpression clones versus parental RH30 cells. Statistics calculated using biological triplicates. *, p < 0.05, **, p < 0.005. (J) Representative images showing the migration of RH30 cells treated with DMSO or Compound 01 across the scratch wound at 0, 24, 48 h post a linear scratch made by a sterile pipette tip. (K) Quantitative bar chart and line chart of the migration area of RH30 cells treated with DMSO or Compound 01. Statistics calculated using biological triplicates. ***, p < 0.0005

Simultaneously, we examined the effects of KDM4B and KDM5A on RMS cell migration, we found that the respective knockouts of KDM4B and KDM5A significantly inhibited migratory capacity (Figs. 3E-F). Overexpression of both targets promoted tumor cell migration phenotypes, further validating the critical roles of KDM4B and KDM5A (Figs. 3G-I). Treatment with Compound 01 also replicated the migration-inhibitory phenotype that was observed following the knockout of two targets, indicating the inhibitory effect of Compound 01 on RMS cells (Figs. 3J-K, Figure S5C).

KDM4B and KDM5A depletion and Compound 01 treatment suppress FP-RMS cell proliferation through apoptosis and cell cycle arrest

Flow cytometry analysis of the two knockout clones showed that the deletion of KDM4B and KDM5A promoted cell apoptosis in fusion-positive RH30 cells (Fig. 4A). Therefore, we measured the apoptosis rate of cells treated with 10 µM Compound 01 for 72 h using Annexin V/PI double staining, showing a significant increase in apoptotic cells, which are consistent with the effect of KDM4B and KDM5A knockout (Fig. 4B). We then performed RNA-seq analysis on RH30 cells treated with 10 µM Compound 01 for 72 h, and the GSEA analysis also enriched apoptosis-related gene sets (Fig. 4C).

Fig. 4.

Compound 01 Mimics KDM4B/KDM5A Loss to Induce Apoptosis and G0/G1 Cell-Cycle Arrest in FP-RMS Cells. (A) KDM4B, KDM5A knockout induced apoptosis compared to parental cells in RH30 determined by flow cytometry. (B) 10 µM cmpound 01 treatment induced apoptosis compared to parental cells in RH30 determined by flow cytometry. (C) Apoptosis pathway enrichment analysis by GSEA for the genes regulated by 10 µM Compound 01 treatment for 72 h of RH30 cells. (D) KDM4B, KDM5A knockout caused G0/G1 arrest in RH30 cells detected by flow cytometry. (E) 10 µM Compound 01 treatment caused G0/G1 and S phases arrest in RH30 cells detected by flow cytometry. (F) GO analysis clustering of commonly differentiated genes in RH30 cells after treatment with 10 µM Compound 01 for 72 h

Additionally, the inhibition of cell proliferation may be related to cell cycle arrest. Flow cytometry analysis showed that both knockout clones induced cell cycle arrest in RH30 cells, indicating their critical roles in regulating cell proliferation. Notably, KDM5A depletion caused a more pronounced reduction in the G2/M phase, suggesting a broader role for KDM5A in regulating both G1-S transition and G2/M progression (Fig. 4D). Treatment with 10 µM Compound 01 also showed a significant increase in the percentage of cells in the G0/G1 phase, indicating that Compound 01 induces G0/G1 phase arrest (Fig. 4E). Gene Ontology (GO) [42] and KEGG enrichment analyses [43] were performed to explore the targeted pathways of Compound 01. The pathways regulated by Compound 01 mainly focused on the negative regulation of cell proliferation and apoptosis pathways (Fig. 4F), this suggests that Compound 01 regulates cell cycle arrest and induces apoptosis in FP-RMS cells through broad-spectrum gene effects.

KDM4B and KDM5A functionally cooperate to sustain the PAX3-FOXO1 core regulatory circuitry

To further confirm the regulatory interactions among KDM4B, KDM5A, and PAX3-FOXO1, we examined the expression of fusion protein in KDM4B and KDM5A knockouts cell lines. The results of immunofluorescence analysis support that the knockout of KDM4B and KDM5A in RH30 cell lines weakened the fluorescence expression of PAX3-FOXO1, indicating a reduction in its expression (Fig. 5A). To investigate the broader effects of KDM4B and KDM5A, RNA sequencing (RNA-seq) was performed to identify transcriptional changes in KDM4B and KDM5A knock-down RH30 cells. GSEA further showed that the differentially down-regulated genes in both knockdown groups were significantly enriched in gene sets related to PAX3-FOXO1 enhancers, indicating that KDM4B and KDM5A knockdown inhibits the activity of key oncogenic transcription factor PAX3-FOXO1 and suppresses its downstream target genes (Figs. 5B-C, Figures S5D-E). Moreover, a significant overlap was observed in the differentially downregulated genes between the KDM4B and KDM5A knockdown groups, suggesting a shared regulatory role of both proteins in gene expression and their complementary actions to some extent (Fig. 5D). A considerable number of PAX3-FOXO1-related target genes and their downstream pathway genes were also common between KDM4B and KDM5A knockdown, supporting the conclusion that KDM4B and KDM5A regulations are highly correlated with PAX3-FOXO1. (Fig. 5D, Figures S5F-G)

Fig. 5.

KDM4B and KDM5A differentially regulate PAX3-FOXO1-driven transcription through RNA-processing and chromatin regulatory complexes. (A) Immunofluorescence analysis of PAX3-FOXO1 expressions in KDM4B and KDM5A knockout clones. Scale bar, 100 μm. (B) GSEA analysis of the differentially down-regulated genes between parental RH30 cells and siRNA-mediated KDM4B-knockdown RH30 cells showing negative enrichment of PAX3-FOXO1 target enhancer-associated genes. (C) GSEA analysis of the differentially down-regulated genes between parental RH30 cells and siRNA-mediated KDM5A-knockdown RH30 cells showing negative enrichment of PAX3-FOXO1 target enhancer-associated genes. (D) Venn diagram for PAX3-FOXO1 target genes and those regulated by KDM4B and KDM5A knockdown in RH30 cells. (E) Functional enrichment of representative PAX3-FOXO1 target pathway genes down-regulated by KDM4B knockdown in RH30 cells. (F) Functional enrichment of representative PAX3-FOXO1 target pathway genes down-regulated by KDM5A knockdown in RH30 cells. (G) Clustering results of immunoprecipitation-mass spectrometry (IP-MS) analysis of KDM4B-interacting proteins. (H) Clustering results of immunoprecipitation-mass spectrometry (IP-MS) analysis of KDM5A-interacting proteins. (I) Western validation of PAX3-FOXO1‘s downstream targets in RH30 and RD cell lines after 72 h treatment with Compound 01. (J) GSEA analysis of the differentially down-regulated genes between RH30 cells treated with Compound 01 or DMSO showing that Compound 01 suppresses genes upregulated in PAX3-FOXO1 fusion-positive rhabdomyosarcoma. (K) GSEA analysis of the differentially down-regulated genes between parental RH30 cells treated with Compound 01 or DMSO showing that Compound 01 inhibits direct target genes of the PAX3-FOXO1 fusion protein. (L) GSEA analysis of the differentially down-regulated genes between parental RH30 cells treated with Compound 01 or DMSO showing that Compound 01 suppresses PAX3-FOXO1 target enhancer-associated genes. (M) Western validation with indicated antibodies in RH30 cell lines after 72 h treatment with Compound 01 showing downregulation of p-AKT and p-MAPK

Functional clustering of these PAX3-FOXO1-related downregulated genes revealed a significant enrichment in transcriptional programs associated with RNA polymerase II-mediated regulation. This suggests that KDM4B and KDM5A depletion leads to the downregulation of transcriptional outputs typically driven by PAX3-FOXO1-mediated recruitment of the RNA Pol II machinery (Figs. 5E-F). Meanwhile, there were differences in downstream effects between KDM4B and KDM5A knockdown: KDM4B knockdown mainly affected RNA splicing and transcription factor DNA binding, whereas KDM5A knockdown was more associated with cell cycle processes and cytoskeletal-related pathways, which may play a role in tumor cell migration and invasion (Figs. 5E-F).

To further elucidate the molecular basis underlying the transcriptional differences between KDM4B and KDM5A knockdown, immunoprecipitation-mass spectrometry (IP-MS) was performed to identify proteins interacting with each demethylase. The interactome of KDM4B was strongly enriched for components of the ribosome and spliceosome (Fig. 5G, Figure S6A), consistent with the RNA-seq observation of KDM4B knockdown preferentially affected pathways. These findings suggest that KDM4B may participate in co-transcriptional or post-transcriptional regulation through its association with RNA processing machinery.

Furthermore, the interactome of KDM5A was enriched for transcriptional regulators and chromatin-associated complexes (Fig. 5H). Several transcription factors closely linked to the PAX3-FOXO1 regulatory network-such as YY1, FOXK1, FOXK2, and MEN1-were detected among KDM5A-associated proteins, suggesting that KDM5A as well as these transcription factors function together within the PAX3-FOXO1 transcriptional regulatory complex (Figure S6B).

We then examined the downstream effect of Compound 01 for its significant reduction of PAX3-FOXO1. Several crucial downstream targets of the fusion protein were detected a reduction in important targets FGFR4 and MYCN in fusion-positive aRMS RH30 cells and a decrease in FGFR4 and MyoD1 in fusion-negative eRMS RD cells (Fig. 5I), indicates that Compound 01 may modulate distinct downstream networks in these two cell types. GSEA revealed that Compound 01 significantly downregulated genes associated with the upregulation of the PAX3-FOXO1 fusion transcription factor, confirming its negative regulation of PAX3-FOXO1 and its target genes (Figs. 5J-K). This indicates that Compound 01 effectively modulates the transcriptional regulatory network associated with PAX3-FOXO1. Additionally, the gene set of downstream target genes regulated by PAX3-FOXO1 enhancers was also significantly negatively enriched (Fig. 5L), suggesting that Compound 01 inhibits the activity of the PAX3-FOXO1 enhancer region, further suppressing its interaction with downstream target genes. Additionally, classic cancer pathways RAS/MAPK, and PI3K/AKT signaling pathways were also affected with a decrease in p-AKT and p-MAPK, suggesting that Compound 01 inhibits these crucial downstream carcinogenic pathways (Fig. 5M).

KDM4B, KDM5A, and PAX3-FOXO1 establish a reciprocal stabilization loop to sustain the super-enhancer-driven transcriptional circuitry

Additionally, we observed that the stability of KDM4B and KDM5A was reciprocally regulated. Specifically, the knockout of KDM4B led to a decrease in KDM5A expression, while the knockout of KDM5A similarly suppressed KDM4B expression (Fig. 6A), suggesting an inter-regulatory relationship between KDM4B and KDM5A. It has been previously validated that the knockdown of KDM4B significantly reduces the expression of PAX3-FOXO1 and its downstream target genes, our study confirmed this finding [22]. We also found that the knockout of KDM5A exerted a similar effect, significantly decreasing the expression of PAX3-FOXO1 and its downstream targets, including FGFR4, MyoD1, and MYCN (Fig. 6B). This indicates that both KDM4B and KDM5A are crucial components of the transcriptional network regulated by PAX3-FOXO1 and may represent vulnerabilities in the oncogenic circuitry of this core transcription factor.

Fig. 6.

KDM4B, KDM5A and PAX3-FOXO1 act as a critical feedback loop and is crucial for PAX3-FOXO1 downstream effects. (A) Western validation of the mutual regulation between KDM4B and KDM5A after knockout of KDM4B or KDM5A in RH30. (B) Western blotting validated the knockout of KDM4B and KDM5A inhibits PAX3-FOXO1 and its downstream targets. (C) Following siRNA-mediated knockdown of KDM4B in RH30 cells, western blotting was performed to detect the expression of KDM5A and PAX3-FOXO1. (D) Real-time quantitative PCR of KDM5A and PAX3-FOXO1 in siRNA-mediated KDM4B-knockdown RH30 cells. **, p < 0.005, ***, p < 0.0005, ns = not significant (p > 0.05). (E) Following siRNA-mediated knockdown of KDM5A in RH30 cells, western blotting was performed to detect the expression of KDM4B and PAX3-FOXO1. (F) Real-time quantitative PCR of KDM4B and PAX3-FOXO1 in siRNA-mediated KDM5A-knockdown RH30 cells. **, p < 0.005, ****, p < 0.00005, ns = not significant (p > 0.05). (G) Following siRNA-mediated knockdown of PAX3-FOXO1 in RH30 cells, western blotting was performed to detect the expression of KDM4B, KDM5A and the downstream targets of fusion protein. (H) Real-time quantitative PCR of KDM4B and KDM5A in siRNA-mediated PAX3-FOXO1-knockdown RH30 cells. **, p < 0.005, ***, p < 0.0005, ns = not significant (p > 0.05). (I) Immunoprecipitation (IP) of RH30 cell lysates was performed using anti-KDM4B antibody, and the immunoprecipitated complex was subjected to Western blotting with the indicated antibodies. (J) Immunoprecipitation (IP) of RH30 cell lysates was performed using anti-KDM5A antibody, and the immunoprecipitated complex was subjected to Western blotting with the indicated antibodies. (K) KDM5A along with KDM4B assists PAX3-FOXO1 in binding DNA and initiating transcription by stabilizing the enhancer complex. KDM4B promotes RNA processing/splicing to ensure the effective expression of oncogenic downstream signaling pathways of PAX3-FOXO1, thus cooperatively sustaining the oncogenic network in RMS

To further confirm the regulatory interactions between KDM4B, KDM5A, and PAX3-FOXO1, we used small interfering RNAs (siRNAs) to specifically target them in FP-RMS cells. The reduction of KDM4B protein in these cells led to a decrease in the expression of both KDM5A and PAX3-FOXO1 (Fig. 6C). qRT-PCR analysis showed that the mRNA levels of KDM5A and PAX3-FOXO1 did not change significantly (Fig. 6D), indicating that this regulation occurs at the post-transcriptional level. Similarly, siRNA targeting KDM5A resulted in a reduction of KDM4B and PAX3-FOXO1 protein expression in RH30 cells (Fig. 6E). While the mRNA levels of KDM4B remained largely unchanged, the mRNA levels of PAX3-FOXO1 increased (Fig. 6F). This suggests that the downregulation of KDM5A may trigger the upregulation of other regulatory factors, which in turn enhance PAX3-FOXO1 mRNA expression. The reduction in PAX3-FOXO1 protein levels may reflect the outcome of complex feedback regulation by KDM5A in RMS cells. Previous studies have shown that KDM4B’s demethylase-dependent and demethylase-independent functions both promote tumor growth [22]. While KDM5A’s role has traditionally been focused on its histone demethylase activity, we sought to confirm whether its effects in this context are attributed to its demethylase function. Using the selective KDM5 demethylase inhibitor CPI-455 [44], we observed that PAX3-FOXO1 expression did not decrease in RH30 cells, and KDM5A expression remained unchanged, despite a significant increase in H3K4Me3 levels under CPI-455 treatment (Figure S7A). These results suggest that KDM5A’s regulation of PAX3-FOXO1 and its downstream signals observed in RMS may be mediated more by a demethylase-independent mechanism.

Finally, siRNA-mediated knockdown of PAX3-FOXO1 also led to a decrease in KDM4B, KDM5A, and the fusion protein’s downstream targets (Fig. 6G). qRT-PCR analysis revealed that knocking down PAX3-FOXO1 did not reduce KDM4B mRNA levels but significantly lowered KDM5A mRNA levels (Fig. 6H), suggesting that PAX3-FOXO1 regulates the expression of these two distinct histone demethylases through different levels of regulation, the effect of PAX3-FOXO1 may be more direct to KDM5A.

The reciprocal regulation of expression indicates a high degree of interdependence between PAX3-FOXO1 and KDM4B, KDM5A. To investigate the physical interactions between PAX3-FOXO1 and KDM4B, KDM5A, we conducted protein-protein interaction assays, as such interactions are known to stabilize protein networks. We confirmed that endogenous KDM4B and KDM5A each form complexes with PAX3-FOXO1 in RH30 cells (Figs. 6I-J).

PAX3-FOXO1 can form super-enhancers at loci of myogenic genes, including MyoD1, which significantly enhances the tumor’s invasive potential by promoting the expression of various oncogenes [9]. Knockdown of PAX3-FOXO1 resulted in a marked reduction in the transcription levels of these downstream oncogenes (Figure S7B). Similarly, knockdown of KDM4B and KDM5A downregulated the activity of PAX3-FOXO1, mimicking the effects of PAX3-FOXO1 knockdown, leading to significant downregulation of its downstream targets (Figures S7C-D). qRT-PCR further revealed differences in how KDM4B and KDM5A regulate individual downstream pathways, suggesting their complementary contributions to regulatory layers of the fusion-driven transcriptional circuitry.

To systematically define the transcriptional consequences of KDM4B and KDM5A depletion, we performed RNA-seq on RH30 cells following siRNA-mediated knockdown. GO analysis revealed that both knockdowns enriched pathways associated with RNA polymerase II-dependent transcriptional regulation, particularly those involving transcriptional activation, DNA-binding transcription factors, and RNA polymerase II-specific initiation factors (Figures S7E-F). These results indicate that KDM4B and KDM5A act as essential cofactors that sustain PAX3-FOXO1-driven enhancer activity, likely by facilitating the assembly or stability of transcription initiation complexes at super-enhancer regions.

The reciprocal protein stabilization with PAX3-FOXO1 is unique to KDM4B and KDM5A among the histone demethylase family members

To assess whether the mutual stabilization loop observed with KDM4B and KDM5A is a unique feature or a general mechanism shared by other histone demethylases in FP-RMS, we expanded our investigation to include other KDM family members. KDM2A, KDM5A, PHF2 and PHF8 were previously identified as PAX3-FOXO1-associated proteins via APEX2-MS analyses [45]; KDM3B was reported to modulate PAX3-FOXO1 oncogenic activity [23]; and KDM5B was selected due to its elevated expression in primary aRMS patient samples [22]. We first evaluated the effect of these candidates on the fusion protein. Knockdown of KDM2A, KDM3B, and KDM5B via siRNA did not significantly affect PAX3-FOXO1 protein levels in RH30 cells (Figure S8A). In contrast, depletion of PHF2 and PHF8 resulted in a marked reduction of PAX3-FOXO1 protein (Figure S8A). Notably, qPCR analysis revealed that the mRNA levels of PAX3-FOXO1 were not decreased following PHF2 or PHF8 knockdown, indicating that PHF2 and PHF8 (Figure S8B), similar to KDM4B and KDM5A, regulate the fusion protein primarily at the post-transcriptional level.

To determine if these interactions form a feedback loop, we examined the reverse regulation. Silencing of PAX3-FOXO1 led to a significant decrease in the mRNA levels of PHF2, suggesting PHF2 is a transcriptional target of the fusion protein (Figure S8B). However, this transcriptional downregulation did not translate to a reduction at the protein level, the protein abundance of PHF2, PHF8, and other tested KDMs remained unchanged following PAX3-FOXO1 depletion (Figure S8A). This discrepancy suggests that while PAX3-FOXO1 promotes the transcription of PHF2, it is not required for maintaining its protein stability.

Collectively, these data demonstrate that only KDM4B and KDM5A are engaged in a distinct, bidirectional positive feedback loop where they mutually maintain protein stability with PAX3-FOXO1.

Discussion

In this study, we identify KDM4B and KDM5A as cooperative epigenetic regulators that sustain the oncogenic transcriptional network driven by the PAX3-FOXO1 fusion protein in fusion-positive rhabdomyosarcoma (FP-RMS). Using a deep learning-based drug discovery strategy, we discovered Compound 01 as a dual-target inhibitor that disrupts both KDM4B and KDM5A function, leading to suppression of PAX3-FOXO1 expression, inhibition of tumor cell proliferation and migration, induction of apoptosis, and cell-cycle arrest. Rather than acting through a single enzymatic pathway, Compound 01 collapses a tightly interconnected epigenetic-transcriptional circuit composed of KDM4B, KDM5A, and PAX3-FOXO1, revealing a previously underappreciated vulnerability in fusion-driven RMS.

Histone demethylases, such as KDM4B and KDM5A, have been increasingly recognized as important epigenetic regulators in cancer [22, 37, 38]. While the role of KDM4B in fusion-driven transcription has been increasingly appreciated, the function of KDM5A in RMS has remained largely unexplored. Notably, KDM5A has been implicated in a broad spectrum of malignancies [41, 46–48], where it functions not only as a histone demethylase but also as a scaffold protein that coordinates transcriptional and chromatin-associated regulatory complexes [49, 50]. Our result demostrate that both KDM4B and KDM5A converge on maintaining PAX3-FOXO1-dependent transcription [22] and functionally cooperate to sustain the PAX3-FOXO1 core regulatory circuitry.

Consistent with this multifaceted role, integrated transcriptomic and proteomic analyses in our study revealed a division of labor between these two demethylases. KDM4B preferentially associates with RNA-processing and spliceosomal machinery and is enriched in pathways related to RNA polymerase II-mediated transcription and RNA splicing, supporting its role in facilitating efficient transcriptional output and RNA maturation. In contrast, KDM5A predominantly interacts with transcription factors and chromatin-associated regulators, including YY1, FOXK1/2, and MEN1-key components of enhancer and super-enhancer assemblies. These interactions align with our observation that KDM5A depletion preferentially affects cell-cycle and cytoskeletal pathways, processes closely linked to tumor proliferation, migration, and invasion.

Compound 01 functionally phenocopies the combined depletion of KDM4B and KDM5A. Treatment with Compound 01 recapitulated the effects observed in genetic knockout models, including reduced proliferation, impaired migration, increased apoptosis, and cell-cycle arrest. At the molecular level, Compound 01 inhibited the demethylase activities of KDM4B and KDM5A, as evidenced by increased H3K9me3 and H3K4me3 levels, respectively, and directly bound the enzymatic pocket of KDM4B through coordination with the active-site metal ion. Importantly, genetic ablation of either KDM4B or KDM5A attenuated the anti-tumor effects of Compound 01, confirming that both proteins serve as essential functional targets of the compound.

Beyond enzymatic inhibition, Compound 01 reduced the protein levels of both KDM4B and KDM5A without altering their mRNA expression, indicating post-transcriptional regulation. This suggests that the potent anti-tumor activity of Compound 01 potentially stems from a combined effect of enzymatic inhibition and protein degradation. While KDM5A degradation was partially rescued by proteasome inhibition, KDM4B appeared to be regulated through a proteasome-independent mechanism, suggesting differential control of protein stability. Notably, KDM4B, KDM5A, and PAX3-FOXO1 exhibited strong reciprocal regulation at the protein level and physically interacted with each other, forming a self-reinforcing feedback loop. Disruption of any component by genetic knockdown or compound treatment destabilized the entire network. Furthermore, selective inhibition of KDM5 demethylase activity alone failed to suppress PAX3-FOXO1 expression, highlighting the importance of demethylase-independent scaffold functions, particularly for KDM5A, in maintaining the fusion-driven transcriptional circuitry.

Collectively, this study establishes KDM4B and KDM5A as central epigenetic nodes that cooperatively sustain the PAX3-FOXO1-driven oncogenic transcriptional program in rhabdomyosarcoma and identifies Compound 01 as a small-molecule disruptor of this circuitry. By simultaneously impairing demethylase activity, protein stability, and transcriptional complex integrity, Compound 01 achieves a multi-layered suppression of fusion-dependent transcription, highlighting the vulnerability of the PAX3-FOXO1 regulatory network to coordinate epigenetic disruption.

Despite these advances, several limitations warrant consideration. Although deep learning-based approaches facilitated efficient identification of candidate KDM inhibitors, the selectivity of Compound 01 toward KDM4B and KDM5A and potential toxicity requires further validation, particularly in light of potential off-target effects. Furthermore, while our study highlights a reciprocal stabilization loop, the potential role of non-histone methylation in regulating PAX3-FOXO1 stability and the precise locus-specific epigenetic landscape at its target enhancers remain to be fully characterized. Addressing these mechanistic nuances, alongside exploring other KDM family members will be essential for refining epigenetic intervention strategies and translating coordinated KDM inhibition into effective clinical therapies.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Author contributions

J.H.Y.: Conceptualization, Performing Experiments, Data Curation, Data Analysis, Original Draft Writing; Q.L.H.: Conceptualization, Supervision; P.X.R.: Supervision; F.B.: Conceptualization, Writing-Review & Editing, Funding Acquisition, Supervision and Project Administration; X.L.Z.: Conceptualization, Writing-Review & Editing, Supervision; K.L.: Conceptualization, Writing-Review & Editing, Funding Acquisition, Supervision and Project Administration.

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

All authors have agreed to the content of the manuscript and agree to this submission.

Competing interests

The authors declare no competing interests.