RBM17 Mediates Evasion of Pro-Leukemic Factors from Splicing-coupled NMD to Enforce Leukemic Stem Cell Maintenance


 Chemo-resistance in acute myeloid leukemia (AML) patients is driven by leukemic stem cells (LSC) resulting in high rates of relapse and low overall survival. Here, we demonstrate that upregulation of the splicing factor, RBM17 preferentially marks and sustains LSCs and directly correlates with shorten patient survival. RBM17 knockdown in primary AML cells leads to myeloid differentiation and impaired colony formation and in vivo engraftment. Integrative multi-omics analyses show that RBM17 repression leads to inclusion of poison exons and production of nonsense-mediated decay (NMD)-sensitive transcripts for pro-leukemic factors and the translation initiation factor, EIF4A2. We show that EIF4A2 is enriched in LSC and its inhibition impairs primary AML progenitor activity. Proteomic analysis of EIF4A2-depleted AML cells shows recapitulation of the RBM17 knockdown biological effects, including pronounced suppression of proteins involved in ribosome biogenesis. Overall, these results provide a rationale to target RBM17 and/or its downstream NMD-sensitive splicing substrates for AML treatment.


Introduction
Acute myeloid leukemia (AML) is a malignant hematopoietic disorder with dysregulated clonal expansion of mutant undifferentiated myeloid progenitor cells, and accounts for approximately 30% of adult leukemias 1 . Despite signi cant advances in cancer therapeutics in recent years, adult AML patients continue to display chemo-resistance at presentation, high relapse rates, and a 5-year overall survival rate less than 25% 1 . AML is maintained by relatively rare populations of leukemic stem cells (LSC) that are responsible for seeding and propagating the disease 2-4 and possess stem cell-like characteristics including the capacity for self-renewal, differentiation potential (albeit limited), and relative quiescence [5][6][7] . This latter property of LSC, as well as their possession of natural resistance mechanisms such as drug e ux pumps, contributes to their intrinsic resistance to conventional chemotherapies that target proliferating cells. In addition, multiple studies have demonstrated that patients whose bulk AML cells have an elevated LSC gene expression signature have worse clinical outcomes 5,8 , suggesting that heightened LSC activity correlates with poor e cacy of conventional therapy.
LSC and other primitive leukemic cells are generally thought to be transformed from hematopoietic stem cells (HSCs) or committed progenitor cells and very often share the same surface markers (CD34 + CD38and CD34 + , respectively) and similar mechanisms that support the self-renewal 9 of their primitive normal counterparts. These similarities make it di cult to speci cally target primitive leukemic cells for drug development. Recently a 78-patient study de ned a panel of 17 LSC signature genes, whose expression levels were shown to be predictive of response to treatment and overall survival for patients treated with daunorubicin and cytarabine 8 . Despite these ndings, there has been limited success in the effort to speci cally target primitive leukemic cells for AML treatment. Therefore, it is essential to gain a more comprehensive understanding of the mechanistic elements that underpin primitive leukemic cell function and that, as such, may represent important and novel therapeutic targets in AML.
Alternative splicing (AS) is one of the major contributors to proteome diversity and is thus tightly controlled throughout normal development 10 . AS is a complex process that involves a variety of regulatory trans-acting splicing factors and responsive cis-acting RNA elements, which act together to determine splice site selection and alternative exon usage 11,12 . Aberrant alternative splicing is recognized as a key driver of cancer, with many of the hallmark processes of cancer being regulated by tumorspeci c splice variants 13 . Dysregulated AS can either alter transcript stability, resulting in changes in protein levels or affect coding potential, leading to expression of proteins with distinctly different functions. In the context of AML, a genome-wide analysis of aberrant AS patterns showed that approximately one third of genes are differentially spliced in the primitive CD34 + cells of AML patients compared to those obtained from normal controls, suggesting that such genes are involved in processes key to cellular function 14 . LSC also have a unique AS pro le when compared to normal aging HSCs, including a switch to pro-survival isoforms, which enhances their maintenance 15 . For example, missplicing of GSK3β enhances the malignant transformation from human pre-leukemic progenitors into self-renewing LSC 16 . Aberrations in AS can result from somatic mutations in splicing factors or in cisacting motifs within exons or introns, or abnormal expression of splicing factors. Analyses of the genomic landscape of AML patients have discovered recurrent mutations in splicing factors SRSF2, SF3B1 and U2AF1, however these mutations are only found in approximately 10% of AML patients studied [17][18][19] . Given that abnormal AS is also prevalent in AML patients with no obvious mutations in RNA splicing genes 14 , it is critical to study the deregulation of splicing factor expression and their underlying mechanism in AML and primitive LSC. In the present study, we focused on aberrant AS in human primitive AML biology and show that RBM17 is preferentially expressed in progenitors and LSC and enacts within these cells an AS program that is critical for supporting their maintenance.

Results
RBM17 expression is associated with primitive AML cells and adverse AML prognosis. Previous studies examining the link between aberrant splicing and AML have focused on spliceosome genes with somatic mutations in AML patients or with abnormal expression levels in bulk AML samples 20,21 . To more broadly pro le splicing factors that may mediate aberrant alternative splicing independently of mutations in AML cells, we performed a data-mining survey of 203 known mRNA splicing factors (members of the "mRNA splicing" and "mRNA alternative splicing" Gene Ontology (GO) categories 22 (Table S1). Strikingly, RNAbinding motif protein 17 (RBM17) was the only splicing factor that was both signi cantly elevated (p=0.008) in the LSC-enriched (LSC+) vs LSC-devoid (LSC-) subsets from 78 karyotypically normal AML patient samples (GSE76008) 8 and strongly linked (p=0.00568) to poor AML prognosis 19 (Figure 1A-C).
We also analyzed the published gene expression pro les of puri ed LT-HSC (Lin -CD34 + CD38 -CD90 + ) from healthy donors and AML samples with normal karyotype (GSE35008) 23 , and observed that RBM17 is expressed at signi cantly higher levels in AML LSC compared to normal LT-HSC ( Figure 1D). Together these ndings suggest that RBM17 is preferentially expressed in primitive AML cells. We went on to validate these results in 8 primary AML samples, along with a unique OCI-AML-8227 AML cell line, which was derived from a primary AML sample and retains an LSC-driven hierarchy 24 . We found that the RBM17 transcript level is signi cantly upregulated in the primitive cell subset (CD34 + ) as compared to the committed cell subset (CD34 -) in OCI-AML-8227 cells ( Figure S1A). In keeping with the increased level of mRNA, RBM17 protein is 1.68 fold higher in the LSC-enriched primitive cell subsets (CD34 + ) of primary AML patient samples (Table S2) compared to the more committed cell subsets (CD34 -) ( Figure 1E, S1B-C), providing further support that elevated RBM17 preferentially marks the primitive compartments of human AML.
To further characterize RBM17 in AML patient AML cells, we analyzed its expression in the LAML-TCGA (https://portal.gdc.cancer.gov/projects/TCGA-LAML) dataset and found that RBM17 expression was signi cantly higher in both poor (n=38) and intermediate (n=76) molecular genetic risk groups compared with the good molecular genetic risk group (n=38) ( Figure 1F). Next, to investigate the gene expression signature of AML patients with high expression of RBM17, we ranked AML patient samples from the GSE76008 dataset based on RBM17 expression level and de ned the top 15% (35 of 221) as RBM17-high cases, and the bottom 15% (35 of 221) as RBM17-low cases. In total, we identi ed 832 differentially expressed genes (FDR≤ 0.05, FC≥2 or ≤ 0.5), including 336 transcripts more abundant, and 496 transcripts less abundant in AML patient samples with higher RBM17 expression ( Figure 1G). Interestingly, of these genes, 82.1% of those co-upregulated with RBM17 in AML are more highly expressed and 73.2% of the genes anti-correlated with RBM17 are expressed at lower levels in LSC+ populations ( Figure 1H), implicating that elevated expression of RBM17 and its associated co-regulated genes overlap with gene expression signature of LSC. To con rm this hypothesis, we carried out Gene Set Enrichment Analysis (GSEA) with the LSC gene set that contains upregulated (UP) and downregulated (DN) genes in LSC 8 , and showed highly signi cant enrichment of the LSC signature gene set in genes coupregulated with RBM17 in AML patients ( Figure 1I). In addition, GSEA also revealed signi cant enrichment of genes involved in ribonucleoprotein complex biogenesis and spliceosomal complex assembly (FDR≤ 0.05) ( Figure 1J) in the set of genes co-upregulated with RBM17. These results together suggest that RBM17 could have an important role in supporting primitive leukemic cell functions.
RBM17 knockdown impairs the stem and progenitor potential of primary AML. To examine the functional roles of RBM17 in primitive AML cells, we knocked down RBM17 using short hairpin RNA (shRNA) in human AML cell lines and patient samples. Brie y, we designed lentiviruses encoding GFP (transduction marker) along with two independent shRNAs, both of which resulted in e cient knockdown of RBM17 after transduction in multiple AML cell lines (shRNA #1 and #2) (Figure 2A, S2A). Interestingly, RBM17 knockdown inhibited AML cell growth ( Figure S2B-D) and induced myeloid differentiation in the HL60 AML cell line ( Figure S2E), the aforementioned OCI-AML-8227 cell line ( Figure S2F), and primary AML cells ( Figure S2G). In addition, RBM17 knockdown in primary AML cells signi cantly reduced colony numbers ( Figure 2B-F, Figure S2H), and impaired survival ( Figure S2I), indicating that RBM17 is required for recognizing AML progenitor potential in vitro. Next, to directly assess the role of RBM17 in LSC growth and survival in vivo, we performed xenograft studies using shRBM17-transduced primary AML specimens. We conducted output/input analysis of the GFP positivity of human (CD45 + ) cells for AML sample #001 ( Figure 2G, S2J) where we achieved a ~35% infection rate, and analysis of the percentage of total CD45 + cells for AML sample #006 where transduction reached saturation (>85%) ( Figure S2K-L). We observed that RBM17 knockdown in these two primary AML samples greatly impeded AML engraftment in transplanted immunode cient mice ( Figure 2H-I). Analysis of the resulting grafts revealed that RBM17 knockdown induced myeloid differentiation in vivo as shown by an increased percentage of mature CD14+ cells in shRBM17 grafts compared to controls ( Figure 2J-K). Together, these ndings indicate RBM17 depletion disrupts primitive AML cell function through enhancing differentiation and inhibiting colony-forming and engraftment capacities. RBM17 controls alternative splicing of genes involved in multiple pathways in AML cells. To understand the molecular mechanisms that underlie the supporting role of the splicing factor RBM17 in AML, we rst performed RBM17 enhanced crosslinking immunoprecipitation (eCLIP)-seq in the K562 myeloid leukemia cell line to identify genome-wide RNA targets bound by RBM17 ( Figure S3A). eCLIP-seq analysis identi ed 866 signi cantly enriched reproducible binding peaks for RBM17 in the genome using a cutoff of FDR<0.05 and log2 (FC)>3 (over size-matched input control), which corresponded to 432 annotated transcripts ( Figure 3A and Table S3). Of these transcripts 93.1% are protein coding genes ( Figure S3B). The majority of these peaks are within the coding sequence (CDS, 20.8%), 5'-splice site (5'-SS, 21.9%), or proximal intronic regions which are closer to intron/exon boundaries (30.4%) ( Figure 3B). The enrichment of RBM17 binding peaks around splice sites is consistent with its known function as a splicing regulator. Through motif analysis, we also identi ed that highly G-enriched motifs mapped to RBM17-binding sites ( Figure 3C). GO analysis of enriched binding sites further showed that transcripts involved in mRNA splicing, RNA processing, translation initiation, DNA repair and protein ubiquitination are preferentially bound by RBM17 ( Figure S3C).
To further identify possible functional consequences of these interactions, we analyzed a published ENCODE RNA-seq dataset of shRBM17 (GSE88633) vs Control (GSE88047) transduced K562 cells 25 . We discovered AS events that are affected by RBM17 knockdown in K562 cells (FDR<0.1, ΔPSI>0.05) (Table  S4), among which exon inclusions of 705 splicing events were supported by RBM17 while the other 633 splicing events were repressed by RBM17 ( Figure 3D). We also observed that RBM17 can be involved in many types of AS events, including Cassette exon (CE), Retained intron (RI), Alternative 3' splice site (A3SS), Alternative 5' splice site (A5SS) and Mutually exclusive exon (MXE), with cassette exons being the most affected ( Figure 3E). We next validated that multiple RBM17-regulated AS events are shared by K562 and HL60 cell lines ( Figure S3D-F). GO analysis revealed that RBM17-affected AS events are involved in a wide range of pathways ranging from RNA processing, cell cycle, intracellular localization, to metabolism and biosynthesis pathways 26 ( Figure 3F).
Integrative multi-omics approaches indicate RBM17-mediated splicing prevents NMD of genes required for leukemic growth. Analogous to genetic mutations, inclusion or exclusion of certain exons or introns can change the reading frame, which would potentially affect conserved regions of the coded protein structure or have deleterious effect on subsequent mRNA translation. To systematically analyze the potential functional links of RBM17-affected splicing events in AML, we applied a published bioinformatics pipeline to predict effects on the corresponding protein upon RBM17 depletion 27 . Through our analysis, we identi ed 88 splicing events yielding changes in "Transcript Biotype" and 70 splicing events resulting in changes in "Protein Domain" through this analysis ( Figure 4A, Table S5). Intriguingly, 13.3% (21/158) of these splicing events cause a complete or partial loss of well annotated protein domains, while a further 32.3% (51/158) of the splicing events are predicted to produce nonsensemediated decay (NMD) sensitive transcripts mainly due to the inclusion of poison exons, and the formation of premature termination codons (PTCs) ( Figure 4A). By overlapping the eCLIP-seq dataset and these 51 NMD sensitive transcripts, we identi ed 6 alternatively spliced transcripts that are predicted to be both bound and regulated by RBM17 ( Figure 4B), suggesting that RBM17 can regulate splicing through direct and speci c binding to the pre-mRNA of these transcripts. Such relationships are in contrast to the previously reported role of RBM17 as a spliceosome component with no preference for its RNA substrate sequences 28 . It is particularly striking that, within these 6 direct splicing targets of RBM17 are EZH2 29,30 , RBM39 20,21,31 and HNRNPDL 32 , all factors that have known important roles in cancer stem cell self-renewal and myeloid malignancy. Together, these studies suggest that RBM17 knockdown leads to NMD of genes involved in leukemia propagation.
To validate whether these RBM17-mediated NMD-sensitive splicing events cause corresponding protein downregulations, we applied LC-MS proteomics to characterize proteome changes after RBM17 knockdown in the K562 cells. At day 5 after lentiviral transduction for RBM17 knockdown, we identi ed 741 proteins with signi cant changes (SigB<0.05) (Table S6). GO analysis showed that these proteins downstream of RBM17 knockdown are enriched in clusters of functional networks representing cell division, RNA processing, autophagy, DNA replication and DNA repair, translation and protein folding and vesicle organization ( Figure 4C). Importantly, when we overlap RBM17-mediated NMD-sensitive splicing targets with the proteomics dataset, among 44 transcripts with measured protein expression values (the other 7 genes were not detected in the proteomics dataset), we demonstrated that 31.8% (17/44) of them are downregulated by RBM17 knockdown at their protein levels (fold change<0.9) ( Figure 4D). GO analysis of these 17 proteins downregulated by RBM17 knockdown through potential NMD further revealed a cluster of RNA processing and RNA splicing genes (Table S7), suggesting that RBM17 regulates a network of RNA-processing proteins involved in RNA homeostasis that both directly and indirectly in uences cancer stem cell biology. We next performed bootstrapping analysis by taking random sets (repeated 10,000 times) of 44 proteins out of the list of 8825 proteins identi ed in our shRBM17 versus shscramble proteomics experiments to calculate the percentage of proteins that were down-regulated (fold change<0.9) within these 10,000 randomly picked 44-protein sets. Strikingly, the median percentage of down regulated proteins from randomly-picked 10,000 runs was 22.2% (the mean percentage is 23.4%), which is signi cantly lower than 31.8% observed for predicted NMD sensitive transcripts (p < 2.2e-6) ( Figure 4E), suggesting that RBM17 indeed controls protein expression through regulating alternative splicing coupled with NMD.
RBM17 suppresses EIF4A2 poison exon inclusion. Our integrative multi-omics RNA-interactome, transcriptome and proteome pro ling analyses revealed that EIF4A2 (eukaryotic translation initiation factor 4A2) is the top direct NMD-sensitive splicing target of RBM17 in leukemic cells ( Figure 5A). RBM17 binds to EIF4A2 intron 10 and RBM17 knockdown promotes the inclusion of a proximal "poison" cassette exon (chr3:186788310-186788416), which is associated with strong depletion of EIF4A2 protein ( Figure  5B). This 107-bp cryptic exon and its anking intronic sequences are highly conserved across vertebrates ( Figure 5C), strongly suggesting that it has a regulatory function 33 . Therefore, we aimed to explore its function as a potential effector of RBM17. We rst validated using isoform-speci c RT-PCR in AML cell lines and primary AML samples that RBM17 knockdown promoted EIF4A2 poison exon inclusion ( Figure  5D-E). Next, to con rm NMD-sensitivity of this EIF4A2 transcript variant that includes the poison exon, we tracked its mRNA decay level after actinomycin D-induced transcription-halt in cells with depletion of UPF1, a protein required for NMD. We demonstrated that the mRNA level of the poison exon-included EIF4A2 variant dropped less with UPF1 knockdown compared to shLuci control ( Figure 5F-G). Conversely, the mRNA level of the poison exon-skipped EIF4A2 variant dropped similarly following UPF1 knockdown as compared to control ( Figure 5H). These results con rmed that RBM17 suppressed the EIF4A2 poison exon inclusion event, which would otherwise trigger EIF4A2 degradation through NMD. Lastly, we validated that RBM17 knockdown in a panel of AML cell lines consistently reduced EIF4A2 protein expression ( Figure 5I). Taken together, our results demonstrate that RBM17 is required for the generation of productive protein-coding transcripts of many pro-leukemic factors and identify EIF4A2 as a bona de direct downstream target of RBM17.
EIF4A2 is elevated in human LSC and is required for leukemogenesis. EIF4A2 encodes an ATP-dependent RNA helicase, which is a subunit of the EIF4F complex involved in ribosome binding to mRNA substrates and scanning for the initiator codon 34 . Intriguingly, through data analysis of the TCGA dataset 35 , we found that EIF4A2 mRNA was more highly expressed in all subtypes of AML than in normal monocytes ( Figure 6A). Importantly, in the context of LSC, just as in the case of RBM17, EIF4A2 is preferentially expressed in LSC-enriched cell fractions compared to LSC-devoid fractions from AML patients at both mRNA and protein levels ( Figure 6B-C). Consistent with RBM17 expression in LT-HSC, EIF4A2 is also more highly expressed in AML cells with a primitive immunophenotype (Lin-CD34 + CD38 -CD90 + ) compared to normal control HSCs ( Figure 6D). Given our demonstration that EIF4A2 is downstream of RBM17, we next aimed to explore the effect of EIF4A2 knockdown on primitive AML cell function. Towards this end, we depleted EIF4A2 in AML cell lines and in primary AML samples using two independent shRNAs (#1 and #2) ( Figure 6E). Strikingly, knockdown of EIF4A2 signi cantly inhibited AML cell growth ( Figure 6F, S4A), induced myeloid differentiation ( Figure 6G-H) and resulted in increased cell apoptosis ( Figure 6I, S4B) as compared to a shscramble control. In addition, depletion of EIF4A2 in three primary AML samples signi cantly inhibited their colony-forming abilities ( Figure 6J-L). These data together indicate that EIF4A2 supports the proliferation, survival and undifferentiated state of AML cells.
EIF4A2 overexpression partially rescues RBM17 knockdown-mediated phenotypes in AML cells. Through correlation analysis, we found that RBM17 supports higher expression of EIF4A2 in two different AML patient datasets ( Figure 7A-B). To test the extent that the downstream effects of RBM17 knockdown are shared upon EIF4A2 knockdown, we rst performed LC-MS proteomics to characterize proteome changes induced by EIF4A2 knockdown in K562 cells. We identi ed a list of signi cantly downregulated and upregulated proteins induced by EIF4A2 knockdown (Table S8). Interestingly, these two gene sets are signi cantly enriched in proteins modulated downstream of RBM17 knockdown as we observed in K562 cells ( Figure 7C-D), suggesting that EIF4A2 knockdown indeed largely recapitulates the biological effects of RBM17 knockdown in AML. Given the signi cant link between EIF4A2 and RBM17 in AML, we next tested whether restoring EIF4A2 could rescue any biological effects caused by RBM17 knockdown. Speci cally, we infected HL60 cells with lentiviruses co-expressing a scramble or RBM17 targeting hairpin with either a luciferase control cDNA or the EIF4A2 cDNA (shscramble+Luci, shscramble+EIF4A2, shRBM17#1+Luci, shRBM17#1+EIF4A2) ( Figure 7E). We found that overexpression of EIF4A2 e ciently reversed the adverse effects of RBM17 knockdown on AML cell apoptosis, and partially rescued AML cell differentiation induced by RBM17 knockdown (Figure 7F, S5A-B).
Interestingly, our subsequent GSEA analysis also revealed that both RBM17 and EIF4A2 knockdown in K562 cells strongly alter expression of proteins enriched in ribosome biogenesis-related gene sets ( Figure   7G-H), and several known translation-related factors ( Figure 7I), indicating that downregulation of RBM17 and EIF4A2 may both affect mRNA translation in leukemic cells. To con rm this nding, we performed an O-propargyl-puromycin (OPP) based protein synthesis assay and detected signi cantly decreased mRNA translation activity in both RBM17-and EIF4A2-knockdown AML cells, respectively ( Figure 7J-K). Moreover, we demonstrated that elevation of EIF4A2 level in RBM17-knockdown AML cells partially rescued the protein synthesis rate ( Figure 7L). These results together suggest that RBM17 inhibits AML apoptosis and differentiation and supports mRNA translation at least partially through enforcing the expression of an NMD-resistant transcript variant and promoting the expression of EIF4A2 protein in human leukemic cells.

Discussion
RBM17, also known as splicing factor 45kDa (SPF45), was originally identi ed as a component of the spliceosome complex. It co-localizes with SR proteins in nuclear speckles and regulates the second step of pre-mRNA splicing by selecting alternative AG splice acceptor sites 36 . RBM17 protein expression is limited in normal tissues and is greatly increased (5-10 fold) in solid tumors of the bladder, lung, colon, breast, ovary, pancreas, and prostate 37 . In addition, RBM17 has been previously linked to cancer chemotherapy resistance in breast and ovarian cancer cell lines through unspeci ed mechanisms 38,39 .
However, the role of RBM17 in AML has not been explored. Interestingly, using proteomics, we previously found that RBM17 is upregulated in human PSCs compared to terminally differentiated broblasts and is required to support PSC self-renewal 40 , a core feature shared in both normal and cancer stem cells. Our work identi ed RBM17 as the sole mRNA splicing factor that is both upregulated in LSC-enriched cell fractions and is signi cantly associated with poor prognosis of AML patients, highlighting the likelihood that RBM17 contributes to the aberrant AS program found in the primitive cells that drive the disease. Through the use of gold-standard in vivo repopulation assays with primary AML samples, we demonstrated that RBM17 depletion impairs the function of disease-and relapse-initiating primitive AML cell compartment. Together these data position RBM17 as a novel leukemic stem cell regulator whose expression and targeting may have important implications in the diagnosis and treatment of malignant hematopoiesis.
A previous study of splicing in mouse neurons showed that RBM17 normally represses the splicing of cryptic junctions and its loss leads to the inclusion of intronic elements in mature transcripts 41 .
Exonization of intronic coding cassettes normally creates frameshifts or introduces PTCs 42,43 . Our integrative multi-omics analysis of RBM17 uncovered, for the rst time, that RBM17 depletion promotes inclusion of poison cassette exons or introns for a number of pro-leukemic factors and leads to their NMD-mediated mRNA degradation and subsequent protein-level downregulation. Our results identi ed the pro-leukemic factors RBM39, EZH2, and HNRNPDL as direct RBM17 mRNA-binding targets with these interactions serving to preserve their protein levels through exclusion of poison exons. A recent study using CRISPR/Cas9 screening demonstrated that complete loss of RBM39 suppresses AML growth both in vitro and in vivo, while pharmacologic RBM39 degradation results in broad anti-leukemic effects 20 . In the same study it was also found that RBM39 loss affects splicing of mRNAs related to RNA-splicing, export, and metablism 20 . Similarly, EZH2 is an important regulator of normal and malignant hematopoiesis 44 , while HNRNPDL overexpression in CML cells has been shown to induce leukemia in vivo 32 . These ndings indicate the possibility that RBM17 knockdown-induced inhibition of these factors contributed to anti-leukemic effects. Our work has therefore provided mechanistic insights into essential AML molecular circuitry by uncovering that the elevated expression of RBM17 serves to selectively represses the formation of PTC containing mRNAs required for supporting LSC function.
Interestingly, our proteomics data showed that RBM17 knockdown causes downregulation of its known spliceosome interactors CHERP and U2SURP (Table S7). This is in line with a previous study in human HEK293T cells showing that RBM17, CHERP and U2SURP reciprocally regulate each other's expression level and share downstream splicing targets enriched for RNA-binding proteins 28 . We speculate that RBM17, and the spliceosome complex it interacts with, collectively block the usage of cryptic splice sites around cassette exons or introns containing PTCs and skip their inclusions. Furthermore, our bootstrapping analysis indicated that RBM17 knockdown leads to downstream gene expression inhibition through splicing-coupled NMD, suggesting that RBM17 preferentially regulates splicing of mRNAs containing PTCs. Exploration of the mechanisms through which RBM17 mediates its speci c splicing of NMD-sensitive transcripts will be of interest to pursue in future studies.
In the present work, we showed that RBM17 represses the inclusion of poison intron10 in the EIF4A2 pre-mRNA, which prevents EIF4A2 mRNA NMD and promotes its downstream protein synthesis. Therefore, EIF4A2 represents a novel potential therapeutics target for AML and intersection between splicing and translation control. Previously, Sadlish et al found that the natural compound rocaglamides stabilizes EIF4A-RNA interactions and interferes with the assembly of the EIF4F complex, thereby blocking translation initiation 45 . More recently, Callahan and colleagues showed that rocaglamide is able to preferentially kill functionally de ned LSC, but relatively spares normal HSPCs through mechanisms beyond simply inhibiting translation initiation 46 . Since rocaglamide does not distinguish EIF4A family members, it has not been clear which member of the EIF4A family underlies the anti-leukemia effects of the compound. Our work clearly demonstrates that EIF4A2 is more highly expressed in LSC and is required to support the proliferation, survival and undifferentiated state of AML cells, indicating the RBM17/EIF4A2 axis we have uncovered is indeed targetable for AML treatment. Importantly, a previous comparative EIF4A1 and EIF4A2 RIP-seq study in HEK293 cells also showed that 23% of EIF4A2's RNA targets are unique 47 . GO analysis of the EIF4A2-speci c RNA targets further showed that these genes are involved in transcription, cell migration, cell cycle and positive regulation of GTPase activity. Many of EIF4A2's unique RNA targets, such as USP6NL (a GTPase-activating protein) and REV1, both of which are signi cantly upregulated in LSC, are indeed downregulated by EIF4A2 knockdown in K562 cells (Table   S7). Thus, understanding the functional role of the EIF4A2-speci c RNA targets in AML and LSC may provide mechanistic support for speci cally targeting EIF4A2 and/or its regulated pathways for AML treatment and directing the improvement of drug target sensitivity.
Upregulation of protein synthesis has been described to occur in "pre-leukemic" myelodysplastic (MDS) stem cells 48 . AML stem cells also exhibit increased expression of ribosome pathway genes 49 , indicating the potential role of ribosome biogenesis in the establishment and propagation of cancer stem cells in the blood system. Importantly, both RBM17 and EIF4A2 knockdown inhibited protein synthesis and downregulated at the protein level, the expression of factors enriched in the ribosome biogenesis pathway, suggesting a link between elevated expression of RBM17 along with its downstream target EIF4A2 and protein synthesis activation in primitive AML cells. Our EIF4A2 rescue experiments support the concept that EIF4A2 inhibition is necessary for the shRBM17-induced apoptosis and contributes to shRBM17-induced myeloid differentiation and translation inhibition. Interestingly, our proteomics analysis showed that RBM17 knockdown also led to downregulation of other translation-related factors including UBA52, EIF4H and EIF3B (Table S6), the collective loss of which may synergize with that of EIF4A2 to further solidify the translation inhibitory effects induced by RBM17 loss.
In conclusion, human AML stem and progenitor cells express abnormally high levels of RBM17 to ultimately enforce NMD-escape for a number of key pro-LSC transcripts (Figure 8). In particular, we place this mechanism of RBM17-directed control upstream of the essential process of protein synthesis in LSC and identify inhibition of the RBM17-EIF4A2 axis as a potential therapeutic avenue for AML treatment. 10ng/ml, Peprotech). OCI-AML-8227 cells were plated 400,000 cells/ 24 well non-adherent culture plate and were split every 6-7 days. Primary AML samples were grown in X-VIVO with 20%BIT, 100 ng/mL human SCF, 100 ng/mL human FLT3, 20 ng/mL human TPO, 20 ng/mL human IL3 and 10 ng/mL human IL3. Human cord blood derived hematopoietic stem and progenitor cells (HPSCs) were cultured in StemSpan SFEM with 20ng/mL human IL-6, 100ng/mL human SCF, 20 ng/mL human TPO and 100ng/mL human FLT3. All cells were incubated at 37℃ in a humidi ed atmosphere containing 5% CO2. All ow cytometry analysis was performed using a BD LSRII ow cytometer, MACSQuant Analyzer and FlowJo Software (v7.6.5). Cell sorting was performed with MoFlo XDP (Beckman Coulter).

AML transduction
For AML cell lines HL60, K562 and NB4, cells were infected with lentivirus using an MOI of 10 followed by puromycin selection or 7-AAD-and GFP+ sorting. For primary AML samples, 1.5 million cells were infected with lentivirus using an MOI of 50 in 24 well ultralow attachment plates with 500ul total growth media, another 500ul of growth media were added 16 hours after infection followed by 7AAD-and GFP+ sorting.
Xenograft studies NSG mice were sublethally irradiated (315 cGy) 1 day prior to injection. Pre-validated engrafting AML sample were infected with shRNA for 24 hours in 24-well culture plates at a multiplicity of infection of 50. Post transduction, cells were validated for GFP expression (monitored sample of cells until 3 days after transduction), washed, resuspended in IMDM + 1% FBS. Then ~200,000 cells were injected with 30ul of IMDM + 1% FBS into the right femur of each recipient mouse, 5 mice were injected per experimental group. 9-12weeks post-transplant, mice were sacri ced and bone marrow from tibias, femurs and pelvis was harvested, crushed with mortar and pestle, ltered and red blood cell lysed using ammonium chloride buffer. Human AML engraftment was analyzed by blocking reconstituted mouse bone marrow with mouse Fc block (BD Biosciences) and human IgG (Sigma), followed by staining with uorochrome-  were obtained from GEO (GSE76008) 8 . Gene expression data on sorted LT-HSC (most primitive hematopoietic cells) from AML patients and healthy controls were obtained from GEO (GSE35008) 23 . Protein expression of xenotransplant validated LSC-enriched and non-LSC fractions from 6 AML samples were obtained from PRIDE (Project PXD008307) 50 . Gene set enrichment analysis (GSEA) was performed by compassion of the "RBM17 high" and "RBM17 low" gene expression pro le with the published LSC gene signature (GSE76008) 8 . RNA-seq data of shRBM17 or Control transduced K562 cells were downloaded from GSE88633 and GSE88047. Data was pre-processed and ltered using standard parameters. Samples were mapped to Human genome (UCSC HG38) with default settings and alternative splicing analyses were performed with rMATS. Functional switches were identi ed using in-house pipelines. Full methodological details are provided in the Supplemental Methods.

Proteomics analyses
Proteomics experiments were performed and analyzed using methods previously described 51 . Full methodological details are provided in the Supplemental Methods.

Statistical Analysis
Sample sizes are indicated in relevant gures. Experiments were repeated at least three times. All statistical analysis was performed using GraphPad Prism (GraphPad Software version 6.0). Unpaired student t-tests were performed with p<0.05 as the cutoff for statistical signi cance. Error bars indicate standard deviation.

Declarations
Data availability The mass spectrometry proteomics raw data have been deposited in the ProteomeXchange Consortium via Proteomics Identi cation (PRIDE) 52 . The accession number of the proteomics data reported in this paper is PRIDE: PXD026780. Publicly available datasets used in this study are described in the Method section of Patient database and RNA-seq analyses. All other data generated during this study are available from the authors on request.       Schematic model depicting the role of RBM17 in primitive AML cells. RBM17 is abnormally higher expressed in the most primitive cell fractions of AML compared to AML blasts, which contributes to e cient splicing of many pro-leukemic factors EZH2, RBM39 and HNRNPDL, along with EIF4A2 that functions in translation control, to sustain LSC functions. Knockdown of RBM17 promotes inclusions of cryptic exons or introns into mRNAs of these pro-leukemic factors, leading to their mRNA degradations due to NMD and consequently resulting in translation blockade, cell apoptosis, limited colony-forming and engraftment capacities, and promoted differentiation in primitive AML cells.