Sequence-dependent splicing dysregulation drives therapy resistance in pediatric AML

Summary

Despite improvements in pediatric acute myeloid leukemia (AML) prognosis, about 30% of patients relapse after initial chemotherapy and have poor survival. However, the genetic basis of resistance remains unclear for most patients. To better understand the mechanistic basis and overcome treatment resistance, we analyze RNA sequencing (RNA-seq) data from 702 pediatric AML patients. This effort uncovers a sequence-dependent splicing dysregulation in 36% of children linked to worse prognosis and a lower rate of complete remission. Surprisingly, this change in RNA splicing matches that induced by SRSF2 mutations, which are common in adult AML. Instead, we identify U2AF2 dysregulation as the driver of aberrant splicing in pediatric AML. The pathologic splicing changes are characterized by “weak” polypyrimidine tracts and are susceptible to modest U2AF2 reduction. These outcomes can be improved by pharmacologic modulation of PRMT enzymes. Overall, these findings highlight the importance of modulating splicing defects to improve treatment response in pediatric AML.

Graphical abstract

Highlights

• •Sequence-dependent splicing defects occur in 36% of pediatric AML
• •U2AF2 downregulation drives aberrant splicing in pediatric AML
• •Weak polypyrimidine tracts increase sensitivity to U2AF2 loss
• •PRMT inhibition partially restores splicing and improves treatment response

Huang et al. show that 36% of pediatric AML patients harbor sequence-dependent splicing defects that resemble those driven by splicing factor mutations in adult AML. They identify U2AF2 downregulation as the key driver linked to treatment resistance and highlight PRMT inhibition as a potential strategy to restore splicing dysregulation.

Introduction

The overall survival (OS) in pediatric patients with acute myeloid leukemia (AML) has been over 65%.¹ However, 30% of patients are refractory to frontline treatment or suffer relapses, which are associated with an extremely poor prognosis.^1,2 While some genetic variations, including somatic mutations on FLT3-ITD, WT1, and KIT, have been recognized as critical drivers of chemoresistance in pediatric AML,^3,4,5 the cause for drug resistance for nearly half of the treatment-refractory patients remains unexplained.⁶

Beyond changes to genomic DNA, there is increasing evidence that chemoresistance is closely linked to transcriptomic dysregulation, particularly changes to the process of alternative mRNA splicing.⁷ Studies have implicated the role of splicing abnormalities in the development of many pediatric cancers. In pediatric B-cell acute lymphoblastic leukemia (B-ALL), hnRNPA1 loss due to abnormal skipping of exon 11 contributes to tumorigenesis by disturbing cell metabolism.⁸ Additionally, aberrant splicing of CD22 drives resistance to CD22-directed immunotherapies in pediatric B-ALL patients.⁹ However, comprehensive explorations of splicing dysregulation in pediatric AML and its possible link to drug resistance are still lacking.

Splicing factors are important proteins that regulate the recognition and processing of splice sites during RNA splicing. Mutations in RNA splicing factors have been identified in various cancers with a strikingly high frequency among adults with myeloid leukemias. For example, nearly 60% of patients with myelodysplastic syndromes,¹⁰ 50% with chronic myelomonocytic leukemia,¹¹ and 30% of adults with AML harbor a somatic mutation in the RNA splicing machinery.¹² Among these, SRSF2 is one of the most commonly mutated genes¹² and leads to widespread splicing abnormalities. Mutations in SRSF2 occur at a specific hotspot residue and lead to widespread splicing abnormalities, which have a characteristic nucleotide preference. For instance, SRSF2 mutations promote the inclusion of an extra exon in the EZH2 gene, introducing premature stop codons that impair hematopoietic differentiation in AML.¹³ However, compared to adult AML, pediatric AML patients rarely harbor mutations in RNA splicing factors,^14,15 and it remains unclear whether splicing dysfunction exists and plays a role in pediatric AML progression and chemoresistance.

To address the above questions, we performed bulk RNA sequencing (RNA-seq) and targeted deep DNA sequencing on 162 pediatric AML patients and included additional 285 RNA-seq data and 255 RNA-seq data from two independent cohorts as validation datasets, constituting the largest collection for splicing research in pediatric AML (n = 702). Despite lacking spliceosomal gene mutations, 36% of pediatric AML patients exhibited pervasive mis-splicing events with previously studied impacts on tumorigenesis. Patients with this splicing abnormality have a poor clinical prognosis and an increased chance of chemoresistance. We further identified reduced expression of the core RNA splicing factor U2AF2 as a key driver of these pathogenic splicing events and chemoresistance. Furthermore, we identify a pharmacologic means to correct this pathologic splicing. Our findings enhanced understanding of cancer-associated splicing dysregulation and suggested a potential strategy to overcome chemoresistance in pediatric AML.

Results

Widespread splicing aberrations in pediatric AML independent of splicing factor mutations

To investigate the pathogenic splicing landscape in pediatric AML, we performed RNA-seq on 162 pediatric AML patients from the Children’s Hospital of Soochow University (Suzhou cohort) (Figure 1A; Table S1). In addition, we collected independent RNA-seq datasets of a Chinese pediatric AML cohort (SCMC, n = 285) and a European pediatric AML cohort (TARGET, n = 255) as validation datasets. We also included an adult AML cohort from The Cancer Genome Atlas ([TCGA] n = 151) to compare the splicing changes between adult and pediatric cases. To rapidly capture the pathogenic splicing changes in pediatric AML, we curated oncogenic splicing events that were previously validated by biochemical experiments and functional studies from publications^16,17,18,19 and the ASCancer Atlas database²⁰ (Table S2). By calculating the presentation and manually checking Integrative Genomics Viewer (IGV), we detected 76 splicing events with varying splicing changes in pediatric AML, among which nearly 90% (68/76) were previously reported to be induced by splicing factor mutations (STAR Methods, Figure 1B). We compared the frequency of the most common splicing factor mutations between our cohort and the TCGA adult AML. As expected, the mutation frequency in adult AML was much higher (17%) than that in our pediatric AML cohort (1%) (Figure S1A; Table S3). Next, by using the PSI (Percent-Splice-In) values of these 76 alternative splicing (AS) events, we performed unsupervised clustering and principal-component analysis on pediatric AML samples of the Suzhou cohort together with 17 CD34⁺ normal controls. This analysis revealed two patient groups with distinct splicing patterns (Figures 1C and S1B). Notably, 36% (59/162) of patients display significantly pervasive occurrences of the 76 aberrant splicing events than the remaining samples (Figures 1C and S1B). We named these two subtypes as “High PSI” and “Low PSI” groups, respectively. Among these 76 splicing events, we found that previously reported pathogenic splicing events such as EZH2 exon inclusion and INTS3 intron retention, known to be driven by SRSF2 mutations in adult AML,^13,17 were consistently observed only in the splicing factor mutant samples in adult AML (Figures 1D and S1C). Remarkably, in contrast, the same events were found in pediatric AML samples lacking splicing factor mutations, which were consistently observed across all three pediatric cohorts (Figures 1D and S1C). In addition, there was a high correlation of the PSI changes between EZH2 exon inclusion and INTS3 intron retention in both pediatric and adult AML cohorts, suggesting a high co-occurrence of these two events in a coordinated manner (Figure S1D).

Figure 1.

Widespread splicing aberrations in pediatric AML independent of splicing factor mutations

(A) Overview of the splicing analysis strategy in pediatric and adult AML.

(B) The reported regulatory factors for each of the 76 splicing events.

(C) Hierarchical clustering and heatmap of the PSI values of the 76 splicing events on 162 pediatric AML samples from the Suzhou cohort and 17 CD34⁺ normal samples.

(D) Read coverage plot showing the inclusion of exon 9b in EZH2 on representative samples of 1 normal sample, 2 adult AML patients, and 6 pediatric AML patients from three cohorts (top). PSI distribution of the EZH2-9b splicing event on all samples in different cohorts, respectively (bottom) (Kruskal-Wallis test).

(E) The overall mis-splicing level in adult (left) and pediatric (right) AMLs ranked by −log(p value). This value represents the overall difference between the PSI values of the 76 splicing events in each patient and the averaged PSI from the Low PSI group (Student’s t test).

(F) RT-PCR results of aberrantly spliced transcripts in pediatric AML patient-derived cells (Suzhou cohort) from the High PSI and Low PSI groups and CD34⁺ cells (n = 2 patients per group).

We next calculated the overall mis-splicing usage of the 76 AS events for each individual by performing a t test between the 76 PSI values in each patient and the averaged 76 PSI values across patients of the Low PSI group as a reference. In adult AML, patients with high aberrant splicing usage were primarily attributed to the presence of splicing factor mutations (Figures 1E and S1E). In contrast, a similar patient distribution was observed in children with AML; however, pediatric AML cases had a very low frequency of splicing factor mutations (Figure 1E). We validated splicing events using isoform-specific reverse-transcription PCR (RT-PCR) in pediatric AML and found that the High PSI group displayed markedly distinct isoform usage from the Low PSI group and normal cells (Figure 1F). To explore whether the observed splicing dysregulation extends beyond the curated 76-gene panel, we performed a genome-wide analysis of AS differences between the High PSI and Low PSI groups. In total, we identified 722 differential splicing events in the High PSI group (Figures S2A–S2C). Functional enrichment analysis revealed that genes were significantly enriched in RNA processing, splicing, and cell cycle pathways (Figure S2D), suggesting a widespread dysregulation of RNA processing networks in High PSI AML. These findings demonstrated that widespread splicing aberrations occur in pediatric AML, which are independent of splicing factor mutation status, suggesting other mechanisms exist to drive these defects.

Pediatric AML patients with splicing dysregulation are associated with poor prognosis and chemoresistance

Given the above splicing heterogeneity in pediatric AML patients, we further explored associations of these changes in splicing with clinical outcomes. Kaplan-Meier survival analysis revealed that patients in the High PSI group experienced significantly shorter OS and event-free survival (EFS) compared to those in the Low PSI group (Figures 2A and 2B). Multivariable CoxPH analysis indicates that the patient subgrouping based on splicing changes is the most significant risk factor for OS and EFS of patients in the Suzhou cohort (Table S4). These survival differences in OS or EFS were validated in two additional pediatric AML cohorts, SCMC and TARGET (Figures 2C–2F). We also examined the relationship between splicing dysregulation and OS in TCGA adult AML and found that patients with splicing factor mutations exhibited worse OS, while this difference was not statistically significant (Figures S1C and S3A).

Figure 2.

Aberrant splicing associated with chemoresistance and poor prognosis in pediatric AML

(A and B) Kaplan-Meier curves for overall survival (A) and event-free survival (B) in pediatric AML patients from the Suzhou cohort (log rank test). Groups were defined by unsupervised clustering of 76 significant splicing events. See STAR Methods.

(C and D) Kaplan-Meier curves for overall survival (C) and event-free survival (D) in pediatric AML patients from the SCMC cohort (log rank test).

(E and F) Kaplan-Meier curves for overall survival (E) and event-free survival (F) in pediatric AML patients from the TARGET cohort (log rank test).

(G and H) Treatment response of first (G) and second (H) induction chemotherapy in pediatric AML from the Suzhou cohort. PR, partial remission; NR, no remission (chi-square test).

(I and J) Treatment response of first (I) and second (J) induction chemotherapy in pediatric AML from the SCMC cohort (chi-square test).

(K and L) Treatment response of first (K) and second (L) induction chemotherapy in pediatric AML from the TARGET cohort (chi-square test).

Besides survival differences, the High PSI group patients exhibited a lower rate of complete remission (CR) following both the first and second rounds of standard induction chemotherapy compared to the Low PSI group (Figures 2G and 2H). This observation was further confirmed in the SCMC and TARGET pediatric AML cohorts (Figures 2I–2L). This strong association between the High PSI group and reduced CR rates underscores the importance of addressing splicing dysregulation as a potential contributor to treatment failure. Furthermore, splicing dysregulation in the High PSI group correlated with a higher residual blast rate following initial chemotherapy and an elevated risk of disease relapse (Figures S3B and S3C). Collectively, these findings suggest a critical role of aberrant splicing in contributing to poorer survival outcomes and increased therapeutic resistance in pediatric AML.

Altered expression of RBPs plays a pivotal role in driving pathogenic splicing changes in pediatric AML

To investigate the underlying cause of aberrant splicing in pediatric AML, we first checked the potential association between genetic mutations and splicing changes. To this end, we employed targeted deep DNA sequencing on 143 matched patients from the Suzhou cohort, whose RNA-seq was used to generate the above splicing analysis (Figure 3A). We identified the somatic mutation landscape of known driver genes and compared it with that from a recently published pediatric AML study.²¹ In general, the mutation frequencies of top driver genes were consistent across the two pediatric AML cohorts (Figure 3B). In the Suzhou cohort, we found that only two patients carried splicing factor mutations (Figure 3A). This limited number is not able to explain the widespread splicing dysregulation observed in 36% of pediatric AML. We, therefore, examined other somatic driver mutations and found similar distributions of these mutations in the High and Low PSI groups. Also, the limited number of mutations cannot explain the highly consistent splicing changes commonly observed in the High PSI group (Figure 3A).

Figure 3.

Exploration of genetic and transcriptomic correlations with aberrant splicing in pediatric AML

(A) Oncoprint displays somatic driver mutations in Suzhou pediatric AML. Cases are divided into High PSI and Low PSI groups, and the genes are grouped based on biological pathways (Fisher’s exact test).

(B) Comparison of mutation frequency between Suzhou and SCMC pediatric AML cohorts.

(C) Heatmap showing the overall gene expression landscape of 380 RBPs in Suzhou pediatric AML. RBP genes are further clustered into three groups based on the Euclidean distance.

(D) The correlations between the 76 AS events and RBP expressions. Edge width means the R values of Mantel’s statistic, and edge color represents the group based on clusters identified in Figures 1C and Figure 3C.

(E) Correlation between RBPs and splicing events. In the left protein interaction network, edge color shows interaction strength (from STRING), node color indicates RBP correlation with the overall splicing scores, and node size reflects the proportion of splicing events containing each binding motif of RBPs. The right heatmap shows the correlation between RBPs and splicing events across patients (Pearson’s correlation). ∗p < 0.05, ∗∗p <0.01, ∗∗∗p < 0.001.

(F) Scatterplot showing U2AF2 gene expression (log₂-normalized FPKM) across groups from the Suzhou pediatric AML (Student’s t test). FPKM: fragments per kilobase of transcript per million mapped reads.

(G) The heatmap shows validated splicing events following each RBP knockdown in the K562 cell line.

Accumulating evidence has highlighted the critical role of altered RNA-binding protein (RBP) expression in driving splicing defects in cancer.^18,22,23 Thus, we next sought to investigate the impact of RBP expression changes on splicing dysregulation in pediatric AML. To this end, we collected a list of human RBPs identified from the RBP2GO database²⁴ and generated the overall gene expression landscape of 380 RBPs between the High and Low PSI groups (Figure 3C). Notably, these RBPs could be further grouped into three distinct clusters based on their varying expression within the High PSI group. Surprisingly, we observed the same pattern embedded in the AS profile within the High PSI group (Figure S3A). More specifically, the similarity between the PSI matrix and RBP expression matrix could be measured by a Mantel matrix correlation coefficient of 0.7 (Figure S4A). This indicates that altered RBP expression may play a role in regulating the splicing changes observed in pediatric AML.

To further identify the splicing targets of the three RBP clusters, we calculated the total correlation between the RBPs of these clusters and each splicing event using the Mantel test. The results indicated that the altered expression of RBPs from cluster C shows the most significant associations with the PSI changes of splicing events identified from the same cluster (Figures 3D and S4A). To pinpoint the key splicing factors at the top of the regulatory hierarchy, we next developed a computational pipeline to identify a core set of RBPs by combining information from RBP-binding motifs, RBP expression correlation with PSI changes, and a ranking-based enrichment analysis (Figure S4C; STAR Methods). From this computational prediction, we identified the top eight candidate RBPs, all belonging to cluster C, and found a dense connection between them by the protein-protein interaction network from the STRING database (Figure 3E). Among them, the RBP U2AF2 (U2 small nuclear RNA auxiliary factor 2) was identified as the top regulator, predicted to modulate splicing and expression of the largest number of the 76 aberrant splicing events (Figure 3E). Significant U2AF2 dysregulation was observed in the High PSI group of the Suzhou pediatric AML patients (Figure 3F).

U2AF2 plays a central role in splice site selection by recognizing 3′ splice sites and recruiting the spliceosome.²⁵ Moreover, we repeated the analysis by using RNA-seq data with RBP knockdown (KD) experiments conducted in K562 leukemia cells from the ENCODE project phase 3 dataset²⁶ and checked the consistency of splicing changes of the 76 events upon each RBP KD. Here again, U2AF2 was validated as the top RBP with the greatest number of regulated splicing events (Figure 3G). In addition, low U2AF2 expression was generally associated with poorer survival and lower chemotherapy response across pediatric AML cohorts (Figure S5). In summary, these findings suggest that in the absence of splicing factor mutations, the pathogenic splicing events in pediatric AML may be primarily regulated by altered RBP expression, with U2AF2 downregulation emerging as a critical regulatory factor.

U2AF2 downregulation causes aberrant splicing events with “weak” polypyrimidine tract

To further validate the impact of U2AF2 on splicing regulation observed in pediatric AML, we performed RNA-seq following U2AF2 KD and overexpression in Kasumi-1 leukemia cells. A total of 1,443 aberrant splicing events induced by U2AF2 KD showed reverse changes upon U2AF2 overexpression (Figures S6A–S6C). Notably, about 86% of the genes aberrantly spliced in our patient cohort were validated by at least one experiment of U2AF2 KD or overexpression (Figure S6D). Genes harboring U2AF2-regulated splicing events were significantly enriched in pathways related to RNA metabolism and splicing, chromatin remodeling, DNA repair, and cell cycle regulation (Figure S6E), providing functional insight into the splicing alterations observed in the High PSI pediatric AML. We further validated several splicing events by RT-PCR in THP1, Kasumi-1, and K562 myeloid leukemia cells (Figures 4 and S7). Consistent with the observations from pediatric AML patients, U2AF2 KD significantly promoted aberrant splicing events, while U2AF2 overexpression reduced expression of these events. Approximately 66% of aberrantly spliced transcripts showing expression changes were reduced in abundance (Figures S6F and S6G) after U2AF2 KD. These findings suggest that U2AF2 dysregulation may also contribute to transcript downregulation by splicing-coupled decay. U2AF2 is known to bind to a uridine or cytidine-rich sequence element upstream of the 3′ splice site, termed the polypyrimidine tract (Py-tract) (Figure 4C). We next used U2AF2 Cross-Linking ImmunoPrecipitation sequence (CLIP-seq) data from the ENCODE dataset²⁶ to explore the transcriptome-wide RNA-binding landscape of U2AF2. Consistent with its role in 3′ splice site selection, U2AF2 binding is highly enriched at the Py-tract region, which also exhibits the most substantial decrease in binding upon U2AF2 KD (Figures 4D and S7G).

Figure 4.

U2AF2 downregulation leads to abnormal splicing events in pediatric AML

(A and B) RT-PCR was performed to analyze mis-splicing events upon U2AF2 knockdown (A) or overexpression (B) in K562 cells (n = 3, mean ± SD; Student’s t test).

(C) Schematic representation of U2AF2 binding locations.

(D) U2AF2-binding sites in WT and KD conditions by CLIP sequencing data from the ENCODE dataset.

(E) INTS3 splicing event and the binding region of the Py-track with U2AF2 were visualized using the CLIP data.

(F) Schematic representation of the INTS3 minigenes with different TC score constructs (left). Splicing analysis of different INTS3 minigenes in K562 cells (right) (n = 3, mean ± SD).

(G) RT-PCR was performed to analyze INTS3 mis-splicing events with different Py-tracks upon U2AF2 knockdown (n = 3, mean ± SD; Student’s t test).

(H) HNRNPH1 splicing event and the binding region of the Py-track with U2AF2 were visualized using the CLIP data.

(I) Schematic representation of the HNRNPH1 minigenes with different TC score constructs (left). Splicing analysis of different HNRNPH1 minigenes in K562 cells (right) (n = 3, mean ± SD).

(J) RT-PCR was performed to analyze HNRNPH1 mis-splicing events with different Py-tracks upon U2AF2 knockdown (n = 3, mean ± SD; Student’s t test).

∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

As the Py-tract is the region for U2AF2 binding, we next sought to investigate the sequence composition of the Py-tract of the 76 splicing events and compare it to that of all introns in the human genome. As expected, the human genome reference sequence demonstrates the stable enrichment of “T” nucleotides 5–30 nucleotides upstream of the 3′ splice sites. In contrast, we observed a sequence variation with a reduced proportion of T nucleotides at the Py-tract of 76 aberrantly spliced events, indicating a “weak” Py-tract²⁷ (Figure S7H). To evaluate how different sequence alterations of the U2AF2 binding site affect splicing, we constructed a series of minigenes that contain sequences from the mis-spliced intron of INTS3 and flanking exons (Figures 4E–4G). CLIP data helped us to identify the wild-type (WT) Py-tract sequence region of the INTS3-retained intron, which is one of the 76 mis-spliced events in pediatric AML (Figure 4E). We then generated 13 variant minigenes with modified Py-tract sequence contents, transitioning from a “strong” tract consisting entirely of T nucleotides to a weak tract composed solely of G nucleotides (Figure 4F). We also assigned a TC score to each modified Py-tract based on the nucleotide content, valuing T at 2 points and C at 1 point. Notably, as the TC score decreased, the intensity of the aberrant splicing events increased gradually (Figure 4F). A similar trend was repeatedly observed in another event with exon skipping of HNRNPH1, which is also one of the 76 aberrantly spliced events (Figures 4H and 4I). Finally, we transfected these minigenes with varying Py-tract strengths into U2AF2 WT and KD leukemia cells to further evaluate the impact of U2AF2 downregulation. In line with previous observations, U2AF2 KD results in elevated utilization of aberrant splicing compared to WT, and more intriguingly, we found that the difference in mis-splicing between KD and WT increases gradually as the Py-tract strength decreases (Figures 4G and 4J). This observation implies that U2AF2 downregulation has limited effects on determining the AS of introns with a strong Py-tract. Conversely, the AS of introns with low TC tract scores, as the 76 oncogenic events we observed in pediatric AML (Figure S7I), appears to be particularly susceptible to modest downregulation of U2AF2 expression.

U2AF2 downregulation leads to chemo-resistance in AML in vitro and vivo

In the above analyses, we found that the High PSI group of pediatric AML patients is associated with poor prognosis and chemotherapy resistance. To further explore this association, we identified 21 cases from the Suzhou cohort characterized by primary chemoresistance (failure of remission-induction therapy) (Table S5). Among these, 13 patients harbor somatic mutations in genes previously reported to be associated with drug resistance, including FLT3, WT1, NRAS, KIT, RUNX1, and TP53. Interestingly, 88% (7/8) of the remaining patients were classified into the High PSI group (Figure 5A), suggesting the potential relationship between splicing abnormalities and resistance phenotype. U2AF2 downregulation was also noted in these patients without known drug resistance mutations (Figure 5B).

Figure 5.

Downregulation of U2AF2 leads to chemotherapy resistance in AML in vitro and vivo

(A) Oncoprint displays mutated genes associated with drug resistance from the Suzhou pediatric AML cohort.

(B) Scatterplot showing U2AF2 expression (log₂-normalized FPKM) in drug-resistant patients from the Suzhou cohort (Student’s t test).

(C) Western blot showing the expression of U2AF2 after U2AF2 knockdown in THP1 and K562 cells.

(D–G) The IC₅₀ value for cytarabine and daunorubicin after U2AF2 knockdown in THP1 and K562 cells. Cell viability was assessed using a CCK8 assay (n = 3; mean ± SD; Student’s t test).

(H) Schematic of Mll-Af9+scramble and Mll-Af9+shU2af2 mouse cells transplantation models in which recipient mice from cohort 1 were euthanized at day 16 post-transplant and mice from cohort 2 were observed for survival analysis until end stage.

(I) Representative photograph of spleens from cohort 1 with an inch scale (left) (n = 5), and spleen weight (right) (n = 5; mean ± SD; Student’s t test). Scale bar, 1 cm. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(J) Representative percentages of GFP+mcherry+ cells of BM and spleen in the recipient from cohort 1 by flow cytometry analysis (n = 5; mean ± SD; Student’s t test). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(K) Kaplan-Meier survival analysis of recipient mice from cohort 2 (n = 5; log rank test).

To investigate the impact of low U2AF2 expression on chemotherapy resistance, we employed small hairpin RNA to suppress U2AF2 expression in THP1, Kasumi-1, and K562 leukemia cells (Figures 5 and S8). We then tested the impact of U2AF2 downregulation on response to standard induction therapy used for pediatric AML, including cytarabine and daunorubicin, and found an increase in the half maximal inhibitory concentration (IC₅₀) value following U2AF2 knockdown (Figures 5D–5G, S8C, and S8D). This observation suggests that U2AF2 downregulation leads to increased resistance to these chemotherapy agents in vitro. Conversely, overexpression of U2AF2 enhanced drug sensitivity and partially reversed chemoresistance to cytarabine and daunorubicin (Figures S8E and S8F).

We next sought to validate this association in vivo by using a transplantation mouse model of myeloid leukemia using Mll-Af9 mouse cells (Figures 5 and S9). C57BL/6 mice engrafted with tumors were injected intravenously with either control scramble cells or shU2af2 mouse cells. The animals implanted with tumors were treated with either control 0.9% NaCl or a combination of cytarabine and daunorubicin on day 11 after transplantation (Figure 5H). Mice with daunorubicin and Ara-C (DA) treatment have reduced spleen size and weight compared to the control group (Figures 5I and S9G). At the same time, the U2af2 KD group showed increased spleen weight compared to the scramble group (Figures 5I and S9G). Additionally, tumor burden was reduced in the daunorubicin/Ara-C treatment groups compared to controls, which was assessed by flow cytometry performed on bone marrow and spleen samples (Figures 5J and S9H). However, the U2af2 KD group exhibited a higher presence of resident tumor cells compared to the scramble group (Figures 5J and S9H). We observed similar leukemia progression in both scrambled and U2af2-KD groups without drug intervention (Figures 5K and S9I). Notably, daunorubicin/Ara-C-treated mice in which AML cells had U2af2 KD exhibited earlier mortality than U2af2 WT counterpart mice (Figures 5K and S9I). This suggests that U2af2 KD confers resistance to induction chemotherapy, resulting in a higher proportion of residual tumor cells after treatment and a shorter OS.

To determine whether aberrant splicing directly affects chemoresistance downstream of U2AF2 dysregulation, we focused on two events predicted to generate premature termination codons in EZH2 and HNRNPH1 (Figure S10A). Both splice isoforms accumulated upon UPF1 KD and showed extended half-lives in actinomycin D chase assays, indicating that they undergo nonsense-mediated decay (NMD) (Figures S10B–S10D). Accordingly, U2AF2 depletion led to reduced EZH2 and HNRNPH1 expression at both mRNA and protein levels (Figures S10E and S10F). Functionally, KD of either EZH2 or HNRNPH1 increased chemoresistance in AML cells (Figures S10G and S10H), suggesting that specific splicing-NMD events contribute to drug resistance following U2AF2 dysregulation.

Combined inhibition of PRMT5 and type I PRMTs rescues aberrant splicing events and decreases chemotherapy resistance

Similar to the commonly mutated spliceosomal proteins, U2AF2 is difficult to target directly due to its essential role in the fundamental RNA splicing process. Previous reports have shown that spliceosomal mutant AML cells are preferentially sensitive to protein arginine methyltransferase (PRMT) inhibition.²⁸ GSK591 and MS023 selectively inhibit PRMT5 and type I PRMTs, respectively. GSK591 suppresses symmetric arginine dimethylation catalyzed by the PRMT5-MEP50 complex, whereas MS023 primarily targets PRMT1-mediated asymmetric dimethylation. The combined inhibition of these enzymes affects spliceosome dynamics and global splicing regulation.²⁸ Considering the potential success of PRMT inhibitors in treating spliceosomal mutant leukemias and the similarity in splicing defects between U2AF2 downregulation and SRSF2 mutation, targeting PRMTs may offer a feasible and effective strategy for treating pediatric AML patients with U2AF2 downregulation.

To investigate the impact of PRMT5 inhibition on AS in pediatric AML, we tested the effect of GSK591 and MS023, both of which have been reported to rescue certain splicing events in adult AML with spliceosomal mutations.²⁸ We collected public RNA-seq data²⁸ from leukemia cells treated with MS023 alone, GSK591 alone, and their combination to assess the effects on rescuing the aberrant splicing changes observed in pediatric AML. The results indicated that 80% of the 76 splicing events could be rescued by either the single or the combined PRMT inhibitors (Figure 6A). Combined inhibitors demonstrated a stronger rescue effect on some aberrant splicing events, such as EZH2, INTS3, and LAS1L (Figure 6B). To validate these findings, we performed RT-PCR, which confirmed the rescue effect of specific aberrant splicing events following treatment with the single inhibitors (Figure 6C) and combined PRMT inhibitors (Figure 6D). We found that some events respond to both inhibitors. For instance, EZH2 exhibits increased aberrant splicing following U2AF2 KD, which can be rescued by GSK591 or MS023. However, some events respond to only one drug. For example, aberrant splicing of CEP164 is effectively suppressed by GSK591 alone, while MKNK1 is rescued only by MS023. We further investigated the effects of combined PRMT inhibitors and found that they effectively rescue aberrant splicing events caused by U2AF2 downregulation. However, the inhibitors have minimal effect in cells with normal U2AF2 expression. These results highlight the increased sensitivity of aberrant splicing to PRMT inhibitors in the context of U2AF2 downregulation. Drug sensitivity assays demonstrated that the combined treatment of Ara-C and MS023/GSK591 significantly reduced the chemotherapy resistance by U2AF2 downregulation (Figure 6E). Overall, our findings suggest that combined PRMT inhibition effectively mitigates the splicing abnormalities and contributes to a decrease in the chemotherapy resistance induced by U2AF2 downregulation.

Figure 6.

Combined inhibition of PRMT5 and type I PRMTs rescues aberrant splicing events and decreases chemotherapy resistance

(A) The heatmap shows the percentage of rescued events among the 76 splicing events in NB4 leukemia cells treated with MS023 alone, GSK591 alone, and a combination of both. “Rescued” events were defined as those showing >1% PSI recovery toward the control level with statistical significance (p < 0.05, paired Student’s t test).

(B) RNA-seq coverage plots of the splicing events in NB4 leukemia cells treated with DMSO, MS023 (MS) alone, GSK591 (GSK) alone, and a combination of both (M + G). The red part in each track below represents aberrant splicing regions in each of the genes.

(C) RT-PCR results detecting aberrantly spliced transcripts in THP1 leukemia cells, which were treated with PRMT inhibitor MS023 or GSK591 with or without U2AF2 knockdown (n = 3; mean ± SD; Student’s t test). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

(D) RT-PCR results detecting aberrantly spliced transcripts in THP1 leukemia cells, which were treated with PRMT inhibitor MS023 and GSK591 combinations with or without U2AF2 knockdown (n = 3; mean ± SD; Student’s t test). ∗p < 0.05, ∗∗p<0.01 ,∗∗∗p < 0.001.

(E) The drug response curve and IC₅₀ value for PRMT inhibitor MS023 and GSK591 combinations after U2AF2 knockdown in THP1 leukemia cells. Cell viability was assessed using a CCK8 assay (n = 3; mean ± SD; Student’s t test).

Discussion

While our study highlights the central role of U2AF2 downregulation in causing splicing defects in pediatric AML, there are additional questions to be addressed for a complete understanding of the detailed regulatory mechanism. First, beyond U2AF2, we observed concomitant changes in the expression of several RBPs in the High PSI group. The dense interconnection between these RBPs suggests an orchestrated regulation of aberrant splicing in pediatric AML. This regulation, in terms of quantitative expression and dynamic interactions between these RBPs, is finely adjusted and functionally selected based on the competitive advantages of tumor cells over other clones. However, replicating the precise expression levels of multiple RBPs simultaneously remains a challenge. We hope that future experimental models and techniques will help to overcome these challenges. Second, the widespread altered expression of RBPs suggests key upstream regulators influencing these proteins. Recent studies have indicated that METTL3 regulates splicing dysregulation by m⁶A-mediated control of splicing factor translation, highlighting a crucial link between m⁶A modification and splicing. Consistent with this, our data indicate that U2AF2 harbors multiple m⁶A modification sites based on methylated RNA immunoprecipitation sequence (MeRIP-seq) analysis (Figure S11A), suggesting it could be a target of METTL3. Indeed, METTL3 KD led to significant downregulation of U2AF2 mRNA expression and reduced U2AF2 protein levels (Figures S11B and S11C), whereas U2AF2 depletion did not affect METTL3 abundance (Figure S11D). These results support a model in which METTL3-mediated m⁶A modification contributes to U2AF2 downregulation and may also broadly influence other RBPs within the splicing network. Further investigation into such epi-transcriptomic regulation may help explain the coordinated RBP dysregulation observed in the High PSI group. Third, in our analysis of the modified Py-tract strength (Figures 4F and 4I), we found that a single nucleotide change in the Py-tract, rather than a gradual PSI change with decreasing TC score, is sufficient to trigger a complete splicing switch. This highlights the need for further investigation into Py-tract sequence determinants for U2AF2 binding and cis-regulation of splicing. Current studies on the splicing impact of individual cis-nucleotide mutations in the Py-tract region are still limited.

The distinct causes of the same mis-splicing events in adult versus pediatric AML suggest that tumors may selectively acquire splicing factor mutations or alter RBP expression to produce the same oncogenic splicing events in a context-dependent manner. It is worth exploring how tumors modulate and switch between these regulatory approaches and what potential factors (i.e., age at diagnosis) influence this selection. One of our hypotheses is that compared to altered RBP expression, which is more feasible and highly dynamic, somatic spliceosomal mutations are more stable DNA changes leading to persistent mis-splicing required for tumorigenesis. However, accumulating such spliceosomal mutations takes time, and these mutations alone are often not sufficient for neoplastic transformation and tumor initiation. In pediatric AML, where there is insufficient time to accumulate such spliceosomal mutations, tumors may resort to altering RBP expressions, a more rapid but less stable way than spliceosomal mutations, to disrupt splicing for their tumorigenesis. Moreover, pediatric hematopoiesis occurs in a distinct developmental and epigenetic context, where progenitor cells are highly sensitive to RNA processing perturbations. Thus, coordinated yet modest RBP expression changes can strongly impact differentiation and survival programs. Unlike adult AML, which harbors mutually exclusive spliceosomal mutations due to synthetic lethality, pediatric AML may achieve splicing dysregulation through flexible, quantitative modulation of RBPs. It is an intriguing mechanism that deserves further investigation.

Our results validated an association between U2AF2 downregulation and chemotherapy resistance. Although U2AF2 downregulation profoundly impacts splicing and treatment response, it does not represent a classical AML driver event. Unlike recurrent splicing factor mutations observed in adult AML, we did not detect inactivating U2AF2 mutations or genomic lesions that confer a clonal growth advantage. Instead, U2AF2 acts as a dosage-sensitive regulator whose reduced expression perturbs splicing and contributes to therapy resistance. Specific mechanisms underlying this resistance need further investigation. It remains unclear which specific splicing events induced by U2AF2 downregulation are responsible for drug resistance. While our cellular data demonstrate that PRMT inhibition can partially restore aberrant splicing and reduce chemoresistance caused by U2AF2 loss, further in vivo validation will be essential to confirm the therapeutic relevance of this approach. Recent preclinical studies have shown that pharmacological inhibition of PRMT5^29,30,31 or type I PRMTs^28,32 effectively suppresses leukemia progression in mouse and patient-derived xenograft (PDX) models by modulating splicing programs and impairing leukemia stem cell maintenance. These findings provide independent support for the translational potential of targeting PRMTs in pediatric AML. Future studies using U2AF2-deficient AML animal models will be instrumental in determining whether PRMT inhibition can specifically reverse U2AF2-associated splicing dysregulation and chemoresistance in vivo. In addition, these effects are not uniquely attributable to U2AF2 modulation. Further investigation into the underlying molecular mechanisms and the development of U2AF2-specific therapeutic strategies may enhance both the precision and efficacy of splicing-targeted interventions. Furthermore, different from adult AML, pediatric AML frequently demonstrates more pronounced epigenetic alterations. However, our exploration of these resistance-driving factors has been limited by the type of sequencing data available, and future studies should aim to integrate and assess the contribution of epigenetic factors to chemotherapy resistance in pediatric AML.

Last, a specific set of AS events has been proposed as a signature to predict treatment response in therapy-resistant leukemia stem cells.^33,34 These AS signatures hold promise as key biomarkers in the clinic, guiding patient stratification for treatment decisions and predicting chemotherapy response. Moreover, splicing signatures could be integrated into standard diagnostics to accurately assess the patient subtyping. With the growing interest in targeting the splicing process for therapeutic purposes, our findings endorse splicing modulation as a promising strategy to overcome chemoresistance in pediatric AML. With a better understanding of molecular mechanisms underlying splicing dysregulation, restoring normal splicing patterns may offer a valuable therapeutic approach to improve outcomes in pediatric AML. In summary, our study highlights the critical role of U2AF2 expression levels in driving splicing dysregulation and contributing to chemotherapy resistance, offering insights into the mis-splicing derived tumorigenesis of this disease. These findings reveal a mechanism of splicing regulation beyond common splicing factor mutations typically found in adult AML and suggest therapeutic strategies to improve treatment outcomes for pediatric AML.

Limitations of the study

This study has several limitations. First, most functional assays were performed in cell line and ex vivo models, which may not fully replicate the developmental and epigenetic context of pediatric hematopoiesis; in vivo validation will be necessary to establish the causal role of U2AF2 downregulation in chemoresistance. Second, although our data suggest METTL3-mediated m⁶A regulation of U2AF2, we did not define the precise methylation sites or the mechanistic sequence linking METTL3 activity to coordinated RBP dysregulation. Third, our analysis of Py-tract determinants is largely associative and additional biochemical or mutational studies will be needed to clarify how single-nucleotide changes drive complete splicing switches. Finally, the lack of integrated epigenomic profiling and in vivo evaluation of PRMT inhibition limits our ability to fully characterize the mechanisms of chemoresistance and assess the therapeutic potential of targeting splicing regulation in pediatric AML.

Resource availability

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Zhaoqi Liu (liuzq@big.ac.cn).

Materials availability

This study did not generate unique reagents.

Data and code availability

• •The raw sequencing data in this study have been deposited in Genome Sequence Archive (HRA002083). The remaining data are available within the article and supplemental information.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This work is supported by the National Key R&D Program of China (grants 2025YFF1207901 and 2023YFF0725400 to Z.L. and grant 2022YFC2502700 to S.H.), the Strategic Priority Research Program of the Chinese Academy of Sciences (grant no. XDA0460403 to Q.W.), the National Natural Science Foundation of China (grant 32170565 to Z.L., grants 82170218 and 82470221 to S.H., and grant 82200177 to L.G.), and the Chinese Academy of Sciences Hundred Talents Program (to Z.L.). We would like to thank the Instrument Sharing Platform of the Chinese Academy of Sciences for providing equipment assistance. We acknowledge all participants involved in the study.

Author contributions

Q.W., Z.L., and S.H. conceived and supervised the entire project. Y.H., L.Q., and L.G. wrote the manuscript with feedback from Q.W., Z.L., S.H., and O.A.-W. Y.H. performed the bioinformatic analyses with the help of L.Z. L.Q. and Y.Z. conducted the experiments with the help of J.S., H.L., and K.C. N.H., X.L., L.Z., and J.S. downloaded and processed public datasets. P.X., L.G., W.H., Y.L., and S.S. performed data collection and clinical information arrangement. All authors read and approved the final manuscript.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
U2AF65 Rabbit polyclonal antibody	Proteintech	Cat#15624-1-AP; RRID:AB_2211330
Anti-GAPDH antibody	Abcam	Cat#Ab9485; RRID:AB_307275
EZH2 rabbit antibody	Cell signaling technogy	Cat#5246; RRID:AB_10694683
HNRNPH1 rabbit antibody	Abclonal	Cat#A5924; RRID:AB_2766663
METTL3 Rabbit Polyclonal Antibody	Proteintech	Cat#15073-1-AP; RRID:AB_2142033
Rabbit IgG HRP linked antibody	Cytiva	Cat#NA934; RRID:AB_772206
FITC Annexin V	Biolegend	Cat#640906; RRID:AB_2561292
Bacterial and virus strains
Pyronic/pcDNA3.1(−)	addgene	Cat#51308; RRID:Addgene_51308
pLKO.1-mCherry-Puro	HonorGene	Cat#HG-VRH1195
pMSCV-IRES-GFP	HonorGene	Cat#HG-VNA0427
Biological samples
Human Acute myeloid leukemia sample	Children’s Hospital of Soochow University	N/A
Human bone marrow transplant donors	Peking University People’s Hospital	N/A
Chemicals, peptides, and recombinant proteins
Cytarabine	MedChemExpress	Cat#HY-13605; CAS#147-94-4
Daunorubicin hydrochloride	MedChemExpress	Cat#HY-13062; CAS#23541-50-6
MS023	MedChemExpress	Cat#HY-19615
GSK591	MedChemExpress	Cat#HY-100235
FuGENE® 4K Transfection Reagent	FuGENE	Cat#4K-1000
Recombinant Murine IL-3	PeproTech	Cat#200-03-100
Recombinant Murine IL-6	PeproTech	Cat#216-16-100
Recombinant Murine GM-CSF	PeproTech	Cat#315-03-100
Murine SCF	PeproTech	Cat#250-03-100
Critical commercial assays
CD34 MicroBead Kit, human	Miltenyi Biotec	Cat#130-046-702
First-strand cDNA Synthesis Mix	Lablead	Cat#F0202
Q5 Site-Directed Mutagenesis Kit	NEB	Cat#E0554S
Cell Counting Kit-8	Lablead	Cat#CK001
QIAamp DNA Blood Mini Kit	QIAGEN	Cat#51104
KAPA RNA HyperPrep Kit	Kapa Biosystems	Cat#KK8540
KAPA RiboErase (HMR) Kit	Kapa Biosystems	Cat#KK8482
EndoFree Plasmid Midi Kit	CWBIO	Cat#CW2105S
KAPA HTP Library Preparation Kit	Kapa Biosystems	Cat#KR0426
Deposited data
RNA-sequencing data	This paper	GSA-Human: HRA002083
Experimental models: Cell lines
Human:THP1 leukemia cell line	G.Huang Laboratory	N/A
Human:K562 leukemia cell line	G.Huang Laboratory	N/A
Human: Kasumi-1 leukemia cell line	X Y. Zhao Laboratory	N/A
Human: MV4-11 leukemia cell line	X Y. Zhao Laboratory	N/A
Human: U937 leukemia cell line	X Y. Zhao Laboratory	N/A
Experimental models: Organisms/strains
C57BL/6N	Charles River	N/A
Oligonucleotides
See Table for a list of human and mouse primer	This paper	N/A
shRNA of scramble: CCTAAGGTTAAGTCGCCCTCG	This paper	N/A
shRNA1 of U2AF2: GAAGAAGAAGGTCCGTAAATA	This paper	N/A
shRNA3 of U2AF2: GGTAGGAACATAGCGTGTTTA	This paper	N/A
Recombinant DNA
U2AF2 cDNA	CUSABIO	Cat#CSB-CL025408HU
METTL3 cDNA	Y G. Yang labratory	N/A
Software and algorithms
GraphPad Prism 10	GraphPad Software Foundation	https://www.graphpad.com
ImageJ	NIH	https://ImageJ.nih.gov/ij/
Samtools (1.9)	Li et al., 200935	https://samtools.sourceforge.net/
BWA (0.7.1)	Li et al., 200936	N/A
STAR (2.7.7a)	Dobin et al., 201337	https://github.com/alexdobin/STAR
featureCounts (2.0.1)	Liao et al., 201438	N/A
Excel	Microsoft	N/A
R software (4.2.3)	R Core Team, 2013	https://www.R-project.org/
Illustrator	Adobe	N/A
rMATS (4.0.2)	Shen et al., 201439	N/A

Experimental model and study participant details

Patient samples

This study was approved by the Medical Ethics Committee of the Children’s Hospital of Soochow University and was conducted by the Declaration of Helsinki. Patients provided samples after informed consent, and primary human de-identified AML samples derived from whole peripheral blood or BM mononuclear cells were used. All samples were anonymously coded according to ethical guidelines. 162 patients were diagnosed and treated at the Children’s Hospital of Soochow University from 2014 to 2021. Normal CD34⁺ cells were derived from bone marrow transplant donors at Peking University People’s Hospital. The raw sequence data reported in this paper have been deposited in the Genome Sequence Archive⁴⁰ in the National Genomics Data Center,⁴¹ China National Center for Bioinformation/Beijing Institute of Genomics, Chinese Academy of Sciences (GSA-Human: HRA002083), which are accessible at https://ngdc.cncb.ac.cn/gsa-human/browse. Controlled accesses to SCMC raw mRNA sequencing data are approved with application number GSA-Human: HRA000789. Controlled accesses to TARGET raw mRNA sequencing data (phs000218.v22.p8) are approved with application project ID: 28370. Controlled accesses to TCGA raw data (phs000178.v11.p8) are approved with application project ID: 35366.

Animals

Animals used in this study were 5- to 6-week-old C57BL/6N healthy male and female mice purchased from Charles River and housed under specific pathogen-free conditions at the animal facility of the Institute of Genetics and Developmental Biology, Chinese Academy of Sciences. All animal care and experiments followed the guidelines of the Institutional Animal Care and Use Committee of the Beijing Institute of Genomics, Chinese Academy of Sciences, Beijing, China. For BM transplantation assays, Mll-Af9 mouse leukemia cells were transfected with scramble or shU2af2 and then selected with puromycin for 3 days. After selection, cells were washed and resuspended in PBS and injected (1 × 106 cells/0.1 mL) into the lateral tail vein of recipient mice. For all mouse experiments, the mice were monitored closely for signs of disease or morbidity daily and were euthanized for survival.

Cell lines and cell culture

THP1 (human acute myeloid leukemia monocytic cell line), Kasumi-1 (human acute myeloid leukemia with t(8;21) chromosomal translocation), and K562 (human chronic myeloid/erythroleukemia cell line) leukemia cells were cultured in RPMI 1640/10% FBS. Mll-Af9 murine leukemia cells were cultured in IMDM/10% FBS plus 10ng/mL mIL-3, mIL-6, mGM-CSF, and mSCF. All cell culture media included penicillin (100 U/mL) and streptomycin (100 μg/mL).

Method details

RT-PCR and quantitative RT-PCR

Total RNA was isolated using TRIzol reagent (Life Sciences) with standard RNA extraction protocol. For cDNA synthesis, total RNA was reverse transcribed with the First-strand cDNA Synthesis Mix (Lablead). Reverse-transcriptase polymerase chain reactions (RT-PCR) were performed on Bio-rad CFX96 or CFX384 PCR cycler using 2xRealab Green PCR Fast mixture (Lablead). The RT-PCR cycling conditions (40 cycles) chosen were as follows: (1) 30 s at 95°C, (2) 10 s at 95°C, (3) 30 s at 60°C with a final 30 s extension at 72°C. Reaction products were resolved on 3% agarose gels and visualized with Super Green staining. For each alternative splicing event, primers flanking the variable exon generated two distinct products corresponding to inclusion and skipping isoforms. Band intensities were quantified using ImageJ after background subtraction. The percent spliced-in (PSI) value was calculated as: PSI = inclusion band intensity/(inclusion + skipping band intensities). This value represents the fraction of transcripts containing the alternative exon/Intron. All reactions were performed in triplicate. Among the 35 significantly altered splicing events upon U2AF2 knockdown, 10 events with ΔPSI >0.05, previously reported to be affected by U2AF1 or SRSF2 mutations, and the normal and aberrant isoforms must have a sufficient and distinct size difference(>40bp) were prioritized. Quantitative RT-PCR was used to assess relative gene expression levels. cDNA was synthesized as described above and amplified using the 2× Realab Green PCR Fast mixture (Lablead) on a Bio-Rad CFX96 or CFX384 instrument. The relative expression levels of target genes were normalized to the housekeeping gene in the same sample using the comparative CT (ΔΔCT) method. Primer sequences are provided in Table S6.

Minigene assay

From K562 genomic DNA, we constructed INTS3-WT and HNRNPH1-WT minigene cloned exons 4 to 5 of human INTS3, exons 4 to 6 of human HNRNPH1 into pcDNA3.1-TFP(−) vector (Invitrogen) using BamHI and XhoI sites, respectively. Artificial mutations were engineered into WT minigene using the Q5 Site-Directed Mutagenesis Kit (NEB) to generate serial mutation minigenes, and the sequences of inserts were verified by Sanger sequencing. Plasmids (1 μg) were transfected using Fugene 4K reagent (Invitrogen) per well of a 12-well plate. Total RNA was extracted 48 h after transfection using TRIzol reagent (Ambion). cDNA was synthesized with the First-strand cDNA Synthesis Mix (Lablead). RT-PCR was done with 2xRealab Green PCR Fast mixture (Lablead) and 40 cycles using primer pairs 5′-TAATACGACTCACTATAGGG-3′ (vector-specific forward primer) and 5′- TCTTCACCAGTTCCCGTACC-3′ (reverse primer for INTS3 minigene), or 5′-CTCCTCTGCCAATGCTGTTA-3′ (reverse primer for HNRNPH1 minigene). Products were run on a 5% PAGE, and the bands were quantified using ImageJ. TFP was used as a control for transfection efficiency and was amplified using a vector-specific forward primer and reverse primer. TFP products were loaded after we ran the INTS3 or HNRNPH1 products for 20–30 min. Percentages of mis-splicing were normalized against exogenous TFP.

Lentivirus or retrovirus infection

Gene overexpression and knockdown were achieved by retroviral or lentiviral transduction, respectively. Lentiviral vector pLKO.1 was used for knockdown, and retroviral vector pMSCV-GFP was used for overexpression. Vectors carried a puromycin resistance gene or GFP for selection, as indicated. The knockdown vector, pLKO.1, containing scrambled shRNA, was used as a negative control. The shRNA against human U2AF2 (hU2AF2), mouse U2af2 (mU2af2), human EZH2 (hEZH2) and human HNRNPH1(hHNRNPH1) were cloned into plKO.1-Puro-Mcherry and/or plKO.1-Puro-Tet-On vector. Transduced cells were selected with puromycin for 3 days. Details of the antisense sequences used are provided in Table S6 (shRNA sequence sheet). For the overexpression of human U2AF2, U2AF2 cDNA sequences were cloned into the retroviral vector pMSCV-GFP, and GFP+ cells were FACS-sorted.

Western blot analysis

Following gel electrophoresis and transfer of proteins to nitrocellulose membranes, the membranes were blocked with Tris-buffered saline (TBS), plus Bovine Serum Albumin (SIGMA) and Tween (5% BSA) (154 mM Tris HCl, 1.37 M NaCl, 0.1% (v/v) Tween 20, 5% (w/v) milk) solution for 1 h. Proteins on membranes were exposed overnight to primary antibody-5% BSA solution. On the following day, membranes were washed in Tris-buffered saline, plus Tween (TBST) (154 mM Tris HCl, 1.37 M NaCl, 0.1% (v/v) Tween 20), for 3 cycles with 10 min wash/incubation for each cycle in TBST to remove unbound antibody. Membranes were further exposed to secondary antibodies in TBST for 60 min. Following secondary antibody incubation, membranes were washed in TBST for 3 cycles with 10 min wash/incubation for each cycle in TBST to remove any unbound secondary antibody. Proteins were detected following a 1-min exposure to Super Signal West Femto Substrate Trial Kit (Pierce). Proteins detected using primary Rabbit anti-human antibodies were: U2AF2 (Proteintech; 15624-1-AP) used at a dilution of 1 in 1000, HNRNPH1 (abclonal; A5924) used at a dilution of 1 in 1000, EZH2 (Cell signaling technology; 5246T) used at a dilution of 1 in 1000, GAPDH (Abcam; ab9485) used at a dilution of 1 in 2000, Secondary antibodies were Rabbit IgG HRP linked antibody(Cytiva; NA934V) used at a dilution of 1 in 5000.

Cell Drug Sensitivity Assay

Cell Counting Kit-8 (Labeled) was used for the Cell Drug Sensitivity Assay. Briefly, dispense 90 μL of cell suspension (10000 cells/well) in a 96-well plate, and add 10 μL of various concentrations of drugs to be tested to the plate. Incubate the plate for 48 h in the incubator. Add 10 μL of the Cell Counting Kit-8 solution to each well of the plate and incubate the plate for 1–4 h. Relative cell viability was determined by measuring the absorbance of each well at 490 nm and converting it to a percent of total viability compared to the untreated control. Drug response assays were conducted either immediately after the puromycin screening (for Figure 6E) or after a short recovery period of 3 days (for Figure 5E).

Targeted DNA sequencing

Genomic DNA was extracted from samples of AML patients using the QIAamp DNA Blood Mini Kit according to the manufacturer’s instructions. The DNA sample was quantified by the Qubit dsDNA BR Assay kit, and DNA integrity was assessed by agarose gel electrophoresis. DNA was sheared on the Covaris M220-focused ultrasonicator. All libraries were prepared using the KAPA HTP Library Preparation Kit according to the manufacturer’s instructions. Fragmented DNA was repaired, 3′ dA-tailed, ligated with Illumina adapters, and size selected, amplified, and assessed using the Agilent 2100 Bioanalyzer. To detect single-nucleotide variation (SNV) and small indels, a customized panel of biotinylated oligoprobes (Roche NimbleGen) was designed to capture all the exons of 1,205 genes that have been identified in previous leukemia sequencing studies. The capture experiment was conducted according to the manufacturer’s protocol. The captured DNA library was finally amplified and sequenced on an Illumina Novoseq 6000 sequencer for 2 × 150 bp paired-end reads.

Somatic mutation calling in targeted DNA sequencing data

Pair-end sequencing FASTQ files were aligned to the human reference genome (GENCODE, GRCh38.99) using BWA (version 0.7.1).³⁶ Samtools (v1.9)³⁵ was used to sort, and index the reads. Somatic mutations and indels were called using SAVI2,⁴² an empirical Bayesian method. Somatic mutations were filtered out if they were annotated as synonymous variants, intragenic variants, or intron variants; annotated as common SNPs in dbSNP138; had a variant allele frequency (VAF) ≤ 5% in the tumor sample; had altered reads ≥2 in the matched normal control; or had an overall mismatch rate >0.02. For tumor-only samples, somatic mutations detected in any normal samples in the cohort or occurring at a frequency >0.01% in population databases (gnomAD, ExAC, 1000 Genomes Project) were additionally filtered out. To annotate and predict the putative pathogenicity of the variants, ANNOVAR’s dbNSFP variant annotation was used.⁴³

Bulk RNA sequencing

Total RNA was extracted from fresh peripheral blood or bone marrow cells of AML samples. For RNA sequencing library preparation, 1 μg of total RNA per sample was processed with the KAPA RNA HyperPrep Kit (Kapa Biosystems, cat. KK8540) following rRNA depletion with the KAPA RiboErase (HMR) Kit (Kapa Biosystems, cat. KK8482). The process included mRNA enrichment, fragmentation, cDNA synthesis, and purification. The library preparation involved end repair, A-tailing, adapter ligation, size selection, and PCR amplification. Library quality was assessed by measuring RNA concentration using a Qubit RNA Assay Kit (Thermofisher), assessing insert size with the Qseq400 system (BIOPTIC) with an R1 cartridge, and quantifying the effective library concentration using the KAPA Library Quantification Kit (KAPA Biosystems). Qualified libraries were pooled and sequenced on the Illumina NovaSeq 6000 platform using a PE150 strategy to generate 150-bp paired-end reads, ensuring comprehensive transcriptome analysis. RNA sequencing data of U2AF2 knockdown and overexpression in Kasumi-1 cells were also generated using the same strategy.

Preprocessing of bulk RNA-seq data

FastQC (v0.11.8) with default parameters was applied to the quality assessment of the raw reads, and all samples passed quality control. The FASTQ sequences were then aligned to the human hg38 reference genome (GENCODE, GRCh38.99) using STAR (v2.7.7a) software³⁷ with default settings. The resulting SAM files were then converted to BAM files, sorted, and indexed using samtools (v1.9).³⁵

Gene expression analysis of bulk RNA-seq data

To generate the mRNA expression matrix for transcriptome analysis, reads were assigned to genes and genic regions using featureCounts (v2.0.1)³⁸ from the Subread package. Genes with low read depths across the cohort were filtered out. Read counts were then transformed into RPKM values, followed by quantile normalization and log₂ transformation at the sample level, as shown in Figures 3F and 5C. Student’s t test was used to test for differential gene expression between sample groups. Dysregulated RBPs were considered to be significant if they passed the p-value threshold of 0.001.

Identification of oncogenic splicing events in pediatric AML

We identified oncogenic splicing events by collecting published papers^16,17,18,19 and previously validated by biochemical experiments and functional studies from the ASCancer Atlas database,²⁰ resulting in a total of 961 splicing event coordinates. Firstly, we selected splicing events with the target gene expressed in the Suzhou pediatric AML cohort. Next, we only consider events with alternative splice sites covered by at least 10 junction reads in the RNA-seq data. These events were further manually checked by using the IGV tool to ensure their presence in the Suzhou pediatric AML cohort and to assess whether splicing variances were observed across patients. In the end, 76 high-confidence splicing events were retained for further analysis. Samples were classified into ‘High PSI’ and ‘Low PSI’ groups by unsupervised hierarchical clustering based on the PSI matrix of 76 significant splicing events using the pheatmap R package. The two clusters exhibited an average PSI difference of 4%, defining the splicing-high and splicing-low subtypes used in subsequent analyses.

Quantification of alternative splicing

Quantification of alternative splicing events was performed using rMATS (v4.0.2)³⁹ using the default parameters. In this study, a “splicing event” refers to a specific alternative exon-intron junction pattern detectable by RNA-seq that differs from the annotated constitutive transcript. Each event is quantified by the percent spliced-in (PSI) value, representing the fraction of transcripts that include the alternative exon or junction (PSI = inclusion/[inclusion + skipping]). RNA-seq datasets of AML patients from the TCGA, TARGET, and SCMC cohorts were used for transcriptome-wide splicing analysis following the same rMATS pipeline. The list of splicing events affected by the knockdown of RBPs (U2AF2, PTBP1, DAZAP1, SF3B4, KHSRP, PCBP1, YBX1, HNRNPA0) was also generated using the same rMATS pipeline. The raw data of the knockdown of RBPs was collected from the ENCODE project phase III dataset.²⁶ The rMATS (v4.0.2) tool was used to perform differential alternative splicing analysis for four major types (SE, A5SS, A3SS, IR) in Kasumi-1 cells after U2AF2 knockdown and overexpression. Splicing events detected with expression in at least 67% (4/6) of samples (read count >30) were selected for analysis. Splicing events with zero Percent Spliced In (PSI) in more than 50% (3/6) of samples were excluded. Significantly differentially spliced events were defined by a p-value <0.05.

Survival analysis

Kaplan-Meier survival curves were created and compared among subgroups for Overall Survival (OS) and Event-Free Survival (EFS) of the patients with the R package ‘Survival ’. The comparison of the OS and EFS between these two groups was performed using the log rank test.

Regulatory RBP analysis

We employed a computational pipeline to identify key splicing regulators among the 380 RBPs (Figure 3C). Firstly, we assessed the correlation between RBP expression and the overall mis-splicing levels across all patients. To quantify the overall mis-splicing levels, For each of these 76 events, we first calculated the mean PSI value across patients in the “Low PSI” group in the Suzhou pediatric AML cohort as the background missplicing rate. Then, for each patient, a paired t test (one-side test toward greater PSI in the tested sample) was applied between PSI of 76 events against the background missplicing rate. Lastly, all patients were ranked based on the ascending order of p values, which indicated the significance level of the abundant missplicing usage of the 76 events. Next, we performed correlation analysis between this patient ranking and RBP expressions and only kept RBPs with an absolute correlation coefficient greater than 0.4 and -log10 p-value >8. Secondly, we applied Gene Set Enrichment Analysis (GSEA) to evaluate the enrichment of aberrant splicing based on RBP-altered expression. For each RBP, we pre-ranked patients according to the RBP expression value and examined the positioning of the patient from the “High PSI” group within this ranking to calculate an enrichment score, retaining RBPs with a normalized enrichment score (NES) less than −3. Thirdly, we calculated the RBP binding motifs from the sequences of the 76 events. The sequence region used for motif analysis is defined by the location of aberrant splicing events, including the two exons and the intervening intron. We gathered motif information from RBPmap,⁴⁴ CisBP,⁴⁵ and MEME⁴⁶ databases, scanning each sequence with FIMO tool⁴⁶ for RBP motifs. We kept RBPs that showed binding motifs in over 70% of the 76 aberrant splicing events. The final list of core regulatory RBPs was generated by overlapping the results from all three steps. This pipeline is illustrated in Figure 3C.

CLIP-seq data processing

Reads were aligned to the human genome (GENCODE, GRCh38.99) using STAR (v2.7.7a), retaining only uniquely mapped reads. To calculate the signal intensity near splice sites, the coordinates for exons 50 bp upstream and introns 300 bp downstream of all 5′ splice sites, as well as introns 300 bp upstream and exons 50 bp downstream of all 3′ splice sites, were first extracted from the reference GTF file (GENCODE, GRCh38.99). The signal intensity at each splice site was then computed using the bedtools. The significant peaks were called using MACS2⁴⁷ with p-value <0.01. Raw sequencing data are available at GEO with accession codes GSE99700 and GSE92143.

Sequence logo analysis

We performed motif analysis at 5′ GT as well as 3′ AG. Motif logo was plotted based on the intron nucleotide composition up to 5 nt upstream and 30 nt downstream of the GT dinucleotide of 5′ss, as well as 30 nt upstream and 5 nt downstream of the AG dinucleotide of 3′ss, using the ggseqlogo packages in Bioconductor.

Calculation of py-track score

We identified 4,017 U2AF2 target genes using CLIP data from U2AF2 KD and WT samples. For each target gene, a 25-bp sequence spanning from 30 bp to 5 bp upstream of each exon was extracted. The results in a number of 159,245 sequences. The number of ‘T’ and ‘C’ bases was counted, with each T assigned 2 points and each C assigned 1 point. Points were summed to calculate the Py-track score for each 25-bp sequence. 159,245 sequences were then ranked by their Py-track scores in ascending order. Finally, the Py-track scores of 76 AS events were analyzed for enrichment using GSEA against this pre-ranked Py-track list as the background signal.

Quantification and statistical analysis

All calculations of clinical data were performed in R statistical software (v4.2.3). For statistical comparison, we performed an unpaired Student’s t test. Kaplan-Meier survival curve p-values were performed using the log rank Mantel-Cox test. The Fisher’s exact test was used to evaluate the relationship between two categorical variables. Correlations were assessed using Spearman’s correlation test. Use the Wilcoxon rank-sum test to identify significant differences between two independent sample groups. The FDR was calculated using the default parameters based on the Benjamin-Hochberg approach. Data with statistical significance are as indicated, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. All experiments were repeated three or more times. Detailed descriptions of experiments and statistical tests were specified in the figure legends.

Published: January 20, 2026