A PML/RARα direct target atlas redefines transcriptional deregulation in acute promyelocytic leukemia

In genetics of myeloid neoplasia, chromosomal translocations often involve transcription factors (TFs), leading to oncogenic fusion TFs that can induce malignant transformation through aberrant transcriptional programs.¹  Depending on the predominant transcriptional activity on direct target genes, these fusion TFs are dichotomously classified as either activators or repressors. This concept of dichotomy has been challenged by the discovery of versatile roles of transcriptional regulators in gene expression,²  largely propelled by rapid advances in high-throughput genomic technologies.

Acute promyelocytic leukemia (APL) is characterized by a specific t(15;17) chromosome translocation, which generates the promyelocytic leukemia/retinoic acid receptor-α (PML/RARα) fusion gene. The encoded fusion protein has been long viewed as a transcriptional repressor of RARα signaling to interfere with myeloid differentiation.³  Our previous studies have revealed that PML/RARα can also exert repressive effects on PU.1 target genes.^4,5  Such repression is largely achieved by the recruitment of corepressors.^3,6  However, gene expression profiling has implicated the capacity to upregulate (in addition to downregulate) gene expression, as observed in both PML/RARα-overexpressed cells and transgenic mice.^7,8  PML/RARα can directly transactivate the expression of genes essential for APL pathogenesis.^9,10  However, only a few such genes have been reported, and the mechanism by which PML/RARα mediates transcriptional activation on a genome scale is unknown.

TF-governing gene regulation is principally modeled as the switch between 2 mutually exclusive classes of coregulators, coactivators and corepressors, respectively activating and repressing transcription.¹¹  However, this classical model has been challenged by genome-wide binding data of coregulators. In human T cells, both coactivators and corepressors can colocalize on actively transcribed regions.²  In embryonic stem cells, the repressor complex and the mediator activator complex are recruited to super-enhancer regions with histone hyperacetylation.¹²  These findings suggest a functional interplay between coactivators and corepressors on chromatin.

Current knowledge of corepressors involved in PML/RARα-mediated transcriptional repression is largely learned from wild-type RARα.^3,13  However, we have shown that RARα-bound canonical retinoic acid responsive elements only account for a small portion of PML/RARα binding sites.⁴  A more interesting observation is that PML/RARα tends to bind open chromatin rather than compact chromatin.¹⁴  These observations strongly suggest that PML/RARα may have distinct coregulator patterns that differ from RARα. However, the limited information regarding direct targets and coregulators has long hampered our understanding of PML/RARα-mediated transcriptional deregulation in APL.

In this study, we established 2 distinct classes of PML/RARα direct targets that were divergent in function during leukemogenesis: repressed and activated genes. Profiling coregulators (HDAC1 and P300) and histone modifications (H3K4me1 and H3K27ac) allowed us to identify super-enhancer-associated PML/RARα-regulated target genes for which expression changes were therapeutically relevant, including the target gene, GFI1.

GFI1 is a transcriptional regulator with context-dependent roles in hematopoiesis and leukemogenesis.^15-23  Reduced or loss of GFI1 expression has been reported in myeloid malignancies,^18-21  while high GFI1 expression has been observed in AML1-ETO-positive acute myeloid leukemia (AML)²³  and acute lymphoid leukemia,²²  but without clear roles defined for GFI1 in APL pathogenesis. Here, we demonstrate that transactivation of GFI1 by PML/RARα (likely oligomerized) and its chromatin conformation regulation at the super-enhancer are necessary for the maintenance of APL. Further profiling GFI1 targets allowed us to identify coregulated genes with PML/RARα.

Primary APL blasts were obtained according to the Declaration of Helsinki at disease onset from bone marrow of 7 newly diagnosed patients with >85% abnormal promyelocytic blasts (supplemental Table 1, available on the Blood Web site). Informed consent was obtained according to procedures approved by the Institutional Review Board from Ruijin Hospital, which is affiliated with Shanghai Jiao Tong University School of Medicine. NB4, U937, and HEK-293T cells were authenticated. Details are available in supplemental Methods.

The integrative assays of chromatin immunoprecipitation sequencing (ChIP-seq) and RNA sequencing (RNA-seq) were performed on NB4 cells without the treatment of all-trans retinoic acid (ATRA) to identify target genes both bound and regulated by PML/RARα. This integration aimed to minimize possible effects associated with each assay, for example, the influence by the changing state of cells upon PML/RARα knockdown. MACS suite²⁴  was used to call ChIP-seq binding peaks (P < 1 × 10⁻¹⁰), and edgeR²⁵  was used to identify differentially expressed genes upon short hairpin RNA (shRNA)-mediated PML/RARα knockdown (false discovery rate <0.01 and at least twofold changes) for 3 days. The maximum distance (50 kb) was used to associate each binding peak with the nearest differential gene showing increased expression (ie, PML/RARα directly repressed targets) or decreased expression (ie, PML/RARα directly activated targets). Details are available in supplemental Methods.

Primers used for these experimental assays are listed in supplemental Table 2, and details are available in supplemental Methods.

Mouse experiments were carried out in accordance with institutional animal protocols, approved by the Institutional Animal Care and Use Committee of Ruijin Hospital/Shanghai Jiao Tong University School of Medicine. The transplantable murine APL mice model was used in the functional studies of Gfi1 and the xenograft nonobese diabetic/severe combined immunodeficiency (NOD/SCID) mouse model used to investigate the effects of the intronic enhancer of GFI1. Details are available in supplemental Methods.

Bioinformatic and statistical analysis are detailed in supplemental Methods.

Identification of bona fide direct targets is essential to investigate the oncogenic activity of PML/RARα during APL pathogenesis. We first sought to identify genes both bound and regulated by PML/RARα in APL cells through integrative assays of ChIP-seq and RNA-seq (Figure 1A). Briefly, we performed ChIP-seq analysis of NB4 cells, an APL patient-derived cell line, using the antibody against the fusion site of PML/RARα (supplemental Figure 1A-C). A total of 6415 PML/RARα binding sites were obtained (supplemental Figure 1D and supplemental Table 3) and confirmed by ChIP-seq data using both anti-RARα and anti-PML antibodies, which were performed by ourselves and retrieved from previously published data²⁶  (supplemental Figure 1E-F). In parallel, we conducted 2 independent RNA-seq experiments on NB4 cells upon PML/RARα knockdown using the fusion gene-specific shRNA (supplemental Figure 2A-C) for 3 days, with a total of 2052 PML/RARα differentially regulated genes identified (supplemental Table 4). PML/RARα knockdown decreased its binding on chromatin (supplemental Figure 2D) and induced differentiation and apoptosis of NB4 cells (supplemental Figure 2E-I). These effects show similarities to those observed when NB4 cells are treated with arsenic trioxide (ATO).^3,27  Combining evidence from ChIP-seq and RNA-seq data, we created an atlas of 787 PML/RARα direct genes, including 363 repressed genes (as expected) and 424 activated genes (unexpected) (Figure 1A; supplemental Table 5). Data from ChIP-quantitative polymerase chain reaction (ChIP-qPCR) using antibodies against the PML moiety, the RARα moiety, or the fusion site of PML/RARα supported the existence of 2 classes of PML/RARα direct targets in APL cells, including both NB4 (supplemental Figure 3) and primary blast cells (Figure 1B). Furthermore, luciferase reporter assays in U937 transfected with PML/RARα and in NB4 cells with PML/RARα knockdown confirmed that transrepression or transactivation was directly mediated by PML/RARα (Figure 1C; supplemental Table 6).

We next investigated the disease relevance of these 2 classes of PML/RARα direct targets. Gene set enrichment analysis (GSEA) using RNA-seq data of 179 AML patients (16 APL and 163 non-APL AML)²⁸  revealed that repressed targets were more likely downregulated in APL, while activated genes tended to be enriched among genes highly expressed in APL (Figure 1D). Similar results were obtained when comparing APL samples and normal promyelocytes²⁹  (Figure 1D), suggesting that both classes of PML/RARα targets were APL specific and malignant related. We also found that these 787 PML/RARα direct genes were collectively informative to distinguish APL patients from non-APL AML patients and normal promyelocytes (supplemental Figure 4). Furthermore, we also analyzed gene signatures for myeloid differentiation stages, including hematopoietic stem cells, myeloid progenitors, and their mature progeny.^30,31  We found that PML/RARα-repressed targets tended to be uniquely expressed in the mature myeloid progeny, ie, band cells and polymorphonuclear neutrophils, consistent with inhibition of terminal neutrophil differentiation by PML/RARα. Interestingly, PML/RARα-activated targets were mostly correlated with genes uniquely expressed in promyelocytes and granulocyte/monocyte progenitors (Figure 1E), suggesting the necessity of PML/RARα in dictating myeloid commitment and the promyelocytic phenotype. We found numerous previously unreported PML/RARα-activated direct targets, including those critical in hematopoiesis and leukemogenesis, such as GFI1, MPO (a strong diagnostic marker for APL³² ), and WT1 and MYC (involved in leukemogenesis^33,34 ).

Next, we explored potential factors determining the PML/RARα-mediated transcriptional consequence (activation vs repression). We speculated that PML/RARα-activated and repressed targets might have different binding patterns of coregulators, considering 2 observations: (1) the recruitment of corepressors is widely recognized to inhibit transcription, whereas the binding of coactivators activates target genes^3,6 ; and (2) the potential of PML/RARα repressing or activating transcription depends on specific contexts (eg, PML/RARα may repress the transcription through recruiting corepressors, and ATRA can reactivate the transcription through replacing corepressors with coactivators).³  To test our speculation, we performed ChIP-seq analysis of corepressor HDAC1 and coactivator P300 in NB4 cells to determine chromatin co-occupancy with PML/RARα (supplemental Tables 7 and 8). We unexpectedly found that both HDAC1 and P300 co-localized with PML/RARα at the majority of PML/RARα binding sites in NB4 cells (Figure 2A). Moreover, we found that binding signals of both coregulators were abundant on the PML/RARα-activated regions, whereas only abundant HDAC1 signals observed on repressed targets (Figure 2B-C). This finding was validated by ChIP-qPCR in primary APL samples (Figure 2D). We also performed co-IP and re-ChIP assays to explore whether HDAC1 and P300 physically bind to PML/RARα and found that the 3 proteins interacted each other on the chromatin of PML/RARα targets (Figure 2E-F).

HDAC1 and P300 are generally assumed to be counterplayers that can control histone acetylation levels determining transcriptional repression or activation.³⁵  The coexistence of HDAC1 and P300 on PML/RARα targets prompted us to further profile the histone acetylation mark H3K27ac. Indeed, the H3K27ac signals on PML/RARα-activated genes were more significantly abundant than on repressed genes (Figure 2G). Histone hyperacetylation is associated with active chromatin.³⁶  We next quantified the degree of active transcription at PML/RARα targets by the traveling ratio, defined as the density of RNA polymerase II (Pol II) at the gene body vs near the promoter.³⁷  Indeed, the traveling ratio was significantly higher on PML/RARα-activated targets than on repressed genes (Figure 2H), indicating that PML/RARα-activated genes were more actively transcribed in APL.

The interaction between PML/RARα and P300 on activated targets suggested the involvement of PML/RARα in enhancer regulation. We thus performed ChIP-seq using the enhancer marks H3K4me1 and H3K27ac to define enhancer regions and found that ∼40% of PML/RARα were located at enhancer regions (Figure 3A). Surprisingly, we also found that PML/RARα preferentially co-occupied enhancers with higher H3K27ac signals (Figure 3B; supplemental Figure 5), which met the criteria of super-enhancers.¹²  We further assessed the binding tendency of PML/RARα on 2 categories of enhancers, super-enhancers and typical enhancers, and found that PML/RARα was over fourfold more likely to co-occupy super-enhancers than typical enhancers (80.6% vs 19.0%; Figure 3C; supplemental Table 9). Similar results (threefold, 60.6% vs 21.5%) were observed in primary APL blasts (Figure 3C; supplemental Table 10). At the gene level, we identified a total of 303 super-enhancer-associated PML/RARα-regulated targets (SEPRTs; supplemental Table 11). To illustrate the functional importance of SEPRTs in APL, we used a CRISPR-Cas9 screening data set in NB4 cells³⁸  to perform GSEA, showing that SEPRTs tended to be enriched among genes essential for cell growth (Figure 3D), such as GFI1, IKZF1, and JUNB.

Next, we explored how SEPRTs responded to ATRA and ATO, 2 widely used drugs in APL therapy. We performed RNA-seq in NB4 cells treated with ATRA or ATO in a time series, from which differentially expressed genes were identified. Using self-organizing map–based gene clustering and visualization, we identified 3 gene clusters (C1-C3; Figure 4A; supplemental Table 12) and found that ATRA and ATO shared some similar regulatory effects (genes in C1 downregulated by both ATRA and ATO) but also exerted different impacts; genes in C2 were only downregulated by ATO, as early as 6 hours after treatment (such as GFI1 and ETS1), while genes in C3 only upregulated by ATRA, with trivial effects upon ATO treatment.

Functional enrichment analysis of gene clusters (Figure 4B; supplemental Table 13) showed that genes in C1 are functionally involved in the apoptotic process and regulation of cell death, such as RUNX2 and CTSH. Genes in C2 are of functional relevance to endocytosis, such as ARRB1 and MAPKAPK3. Genes in C3 are functionally essential for myeloid leukocyte activation (eg, SPI1, SUCNR1, and PIK3CD), suggesting that these genes might contribute to ATRA-induced differentiation. Also, genes associated with coagulation (eg, PLEK and PTPN6) were enriched in C3, suggesting that the effect of ATRA in alleviating coagulation may be attributed to the regulation of SEPRTs as well. We also explored evidence supporting therapeutic potential by identifying the crosstalk between pathways from our time-course expression data and pathway interaction data (see supplemental Methods). This identified synergistically regulated genes (CTSH and PRKCD) and targets (PIK3CD and RRM2) of approved drugs in other diseases, supporting repurposing opportunities and potential efficacy (Figure 4C). Together, these results indicated that ATO and ATRA can act through SEPRTs, likely explaining the combination use in APL therapy.

The functional significance of PML/RARα-mediated repression in APL has been intensively investigated; however, little attention is paid to PML/RARα-mediated activation. As an exemplar, we performed more detailed functional experiments on GFI1, the gene encoding a TF known to be involved in the development of hematopoietic stem cells³⁹  and lymphocytes.²²  We found that GFI1 knockdown induced a partial granulocytic differentiation and apoptosis, promoted the G0/G1 phase arrest, and reduced the colony-forming abilities of NB4 cells (Figure 5A-E; supplemental Figure 6A). This finding was also verified in primary APL blasts (Figure 5F; supplemental Figure 6B).

Moreover, we determined the in vivo role of Gfi1 using a transplantable APL mouse model, which closely mimics human APL with similar biological characteristics.⁴⁰  PML/RARα-positive leukemic cells from APL transplantable mice were transduced with shRNAs targeting Gfi1 and then transplanted into the recipient FVB/NJ mice (Figure 5G; supplemental Figure 7). As shown in Figure 5H-K, transplantable blasts with Gfi1 knockdown could not efficiently develop leukemia, as characterized by no obvious tumor burden observed in the recipients. The results suggested that GFI1 is necessary for the maintenance of APL.

Since PML/RARα can form as heterodimers or oligomers through the RBCC domain,⁴¹  it is tempting to speculate that PML/RARα might transactivate GFI1 through chromatin conformation at the super-enhancer region. We first exploited 3C-PCR⁴²  to examine the chromatin structure at the super-enhancer region of GFI1 in NB4 cells (Figure 6A). There existed 3 PML/RARα-bound cis elements at the GFI1 locus: the promoter (peak 1), the intronic enhancer (peak 2) and the distal enhancer (peak 3). The loop linking the intronic enhancer (peak 2) to the other 2 cis elements showed the strongest interaction, consistent with the highest PML/RARα binding on this site observed in ChIP-seq (top-left panel in Figure 6A).

Using a PML/RARα-inducible cell line (supplemental Figure 8A), we found that the interaction between the intronic enhancer and the promoter became significantly stronger upon PML/RARα induction (Figure 6B), confirming the importance of PML/RARα for super-enhancer formation, mainly through binding on the intronic enhancer. To test the outcome of PML/RARα-mediated chromatin conformation, we generated a construct containing the intronic enhancer cloned upstream of the GFI1 promoter and found that this construct resulted in a 2.5-fold increase in luciferase activity compared with the promoter construct lacking the enhancer activity (Figure 6C). We further used PML/RARα-F158E (with the mutation in the RBCC domain impairing oligomerization⁴³ ) to determine the contribution of PML/RARα oligomerization. We found that chromatin conformation at the GFI1 locus was barely seen and the associated transcriptional activation was significantly reduced in mutant cells (Figure 6D-E; supplemental Figure 8B), indicating that PML/RARα, likely oligomerized, transactivated GFI1 through chromatin conformation.

Next, we determined the effect of PML/RARα binding on super-enhancer regulation in APL development using CRIPPR-dCas9-mediated technology,⁴⁴  which is capable of disrupting PML/RARα binding and eliciting a silencing event on the targeted region. We designed the single guide RNA (sgRNA) targeting the PML/RARα-enriched loci within the intronic enhancer region of GFI1 (sg-GFI1e) and transduced it into dCas9-KRAB stably expressing NB4 cells (Figure 6F). As shown in Figure 6G, the perturbation of PML/RARα binding led to the reduced GFI1 expression, demonstrating that GFI1 activation depended on PML/RARα binding to this enhancer. Also, these sgRNA-transduced dCas9-KRAB-expressing cells were able to differentiate into terminal granulocytes (Figure 6G; supplemental Figure 9). When these cells transplanted into NOD/SCID mice, we found that leukemia could not be efficiently developed (Figure 6H). Together, the in vitro and in vivo evidence demonstrated that PML/RARα-mediated activation on the GFI1 super-enhancer was required for the development of APL.

We reasoned that analyzing GFI1 targets particularly comparing with PML/RARα targets could allow us to deduce the interplay between GFI1 and PML/RARα in APL cells. We first performed ChIP-seq of GFI1 in NB4 cells and identified 1603 binding sites shared by both PML/RARα and GFI1 (Figure 7A; supplemental Table 14). Interestingly, motif analysis revealed that the most significant motifs were those for myeloid TFs (the top 3 included CEBPA, ETS, and RUNX1) (Figure 7B), whereas for all GFI1 ChIP regions, the most significant was as expected the canonical GFI1 motif (supplemental Figure 10A). This observation was supported by published ChIP-seq data on RUNX1⁴⁵  and the ETS family member FLI1⁴⁶  (Figure 7C; supplemental Figure 10B). This suggested that PML/RARα and GFI1 might be assembled with a number of myeloid TFs to coregulate the corresponding target genes, consistent with the notion that multiple TFs coordinately regulate lineage-specific gene expression.⁴⁷  Functional enrichment analysis supported that coregulated target genes were indeed of functional relevance to myeloid differentiation and myeloid-related activities (Figure 7D).

Next, we conducted RNA-seq experiments in NB4 cells with and without GFI1 knockdown. GSEA showed that coregulated genes were significantly enriched within GFI1-repressed genes (Figure 7E), consistent with the previously identified role of GFI1 as a transcriptional repressor.⁴⁸  Interestingly, coregulated genes were also significantly enriched within GFI1-activated genes (Figure 7E). When inspecting genes significantly regulated upon PML/RARα knockdown, we found that PML/RARα also activated these GFI1-activated genes, such as CXCR4 and MPO (supplemental Figure 10C). This motivated us to further explore the role of GFI1 on PML/RARα-mediated transactivation on these coregulated genes. We found that the binding of PML/RARα on coactivated genes was dramatically decreased in NB4 cells after GFI1 knockdown (Figure 7F). Together, the results suggested that GFI1 might coordinate with the activating function of PML/RARα on these cobound and coactivated genes.

Genome-wide identification of direct targets, particularly a larger number of activated targets, challenges the longstanding view of PML/RARα being a transcriptional repressor. This direct target atlas provides a more complete picture of PML/RARα-driven oncogenesis: PML/RARα represses genes responsible for hematopoiesis and activates super-enhancer genes responsible for APL-specific characteristics.

We redefine the activation role of PML/RARα in APL. Particularly, we provide rich evidence supporting the functional importance of PML/RARα-activated target gene GFI1 in APL. We demonstrate that GFI1 can coordinate with PML/RARα to maintain APL and that PML/RARα-induced differentiation block can be relieved upon GFI1 knockdown. Our findings differ from the previous reports that loss or reduced level of GFI1 impedes generation of granulocytes^49,50  but rather support that GFI1 has context-dependent roles in leukemogenesis (in APL depending on PML/RARα). Of note, by reanalyzing the expression data from APL blasts and normal promyelocytes,²⁹  we reveal that GFI1 is maintained at the high expression level in both samples, though PML/RARα-activated targets as a whole tend to be malignant related (APL vs normal promyelocytes) and APL specific (APL vs non-APL AML) (Figure 1D; supplemental Figure 4). It implicates that some of activated targets, such as GFI1, are dedicated to transcriptional programs needed to maintain the promyelocytic stage for APL (maintained consistently) or normal promyelocytes (transiently). We have observed very little overlap of our direct targets with differentially expressed genes reported from the PML/RARα-transgenic preleukemic mice study, in keeping with previous studies.²³  These observations call for future studies to clarify the exact function of GFI1 in the initiation and development of APL as well as promyelocyte development.

Equally interesting is that ATO can significantly reverse the activation of GFI1, whereas ATRA has minimal effect (Figure 4A). This seems to be consistent with (1) our observation that ATO, capable of rapidly degrading PML/RARα,^41,51  can dramatically reduce PML/RARα, HDAC1, and P300 binding and also interrupt the chromatin conformation at the super-enhancer region of GFI1 (supplemental Figure 11); and (2) the fact that ATRA can release PML/RARα binding only at the late stage of differentiation by activating the ubiquitin proteasome system.^52,53 

How PML/RARα activates gene expression has been largely undocumented. The prevailing view is that before ATRA treatment, PML/RARα recruits only HDAC1, which is then replaced with P300 after ATRA treatment.³  Our data challenges this view. We show that in the absence of retinoic acid, PML/RARα can interact with abundant P300 and HDAC1 on chromatin (though not necessarily colocalized on the same site of chromatin) to activate gene expression, likely in a ligand-independent manner. Although fetal bovine serum used for cell culture might contain traces of retinoids, the contribution of ATRA might be below the physiological level (supplemental Figure 12). The further addition of the physiological concentration (1 nM) of ATRA has trivial impact on such activation (supplemental Figure 12), supporting ligand independency of PML/RARα-mediated activation. There is considerable evidence to support our findings. Firstly, histone deacetylases (HDACs) may help prevent inappropriate reinitiation of transcription on regulatory regions simultaneously bound by histone acetyltransferases and HDACs.⁵⁴  Secondly, genome-wide profiling of chromatin-modifying enzymes has indicated that histone acetyltransferases and HDACs can be colocalized on highly active transcribed regions.^2,55  Thirdly, histone deacetylases can also act as coactivators at a variety of genes.^2,56-58  The tendency of PML/RARα binding on open chromatin rather than compact chromatin¹⁴  strongly supports our findings. PML has been reported to exist in the PML/RARα heterodimers/oligomers,^59-61  and P300 can interact with wild-type PML or PML/RARα.^62,63  It is deducible that P300 might interact with PML/RARα either through PML within the PML-PML/RARα oligomer or through the PML moiety of PML/RARα. Whether an oligomerized PML/RARα forms a complex with both P300 and HDAC1 simultaneously requires more evidence of future studies to support.

Inaddition to the newly identified role of transcriptional activation, we have also found that PML/RARα, likely oligomerized, activates GFI1 through chromatin conformation at the super-enhancer region. This provides the structural basis for the efficient transcriptional activation. PML/RARα could form oligomer via the RBCC domain,^43,64  and this property enables PML/RARα to interact with each other to regulate targets through the chromatin conformation in APL.

We note some limitations to our work, which was largely limited by available techniques. First, PML/RARα target genes are identified through integrating RNA-seq (shRNA knockdown) with ChIP-seq (against fusion site); such integration maximizes the power of each technique but cannot distinguish the regulation in a manner whether or not depending on DNA binding. Second, biochemical mechanisms of PML/RARα action on chromatin looping to gene activation await investigation. Third, high-throughput assays are much needed to extensively validate PML/RARα targets, especially activated targets. We anticipate that our work can be further extended with future technological innovations.

Finally, it is worth mentioning that small-molecule compounds selectively targeting super-enhancers have shown tremendous therapeutic potential in inhibiting cancer cell growth.⁶⁵  Inhibition of PML/RARα-regulated super-enhancers, such as targeting GFI1 demonstrated in this study, may be one of the promising mechanisms that can be exploited pharmacologically to extend the success of ATRA/ATO therapy in APL.

The authors gratefully acknowledge the invaluable comments from Hugues de Thé (Université de Paris 7 Denis Diderot, Paris, France), the Center for High-Performance Computing at Shanghai Jiao Tong University for providing the computing platform, and Xujie Zhao for his helpful early inputs on this project.

H.F. is supported by the Program for Professor of Special Appointment (Eastern Scholar) at Shanghai Institutions of Higher Learning. This work was supported, in part, by the National Natural Science Foundation of China (grants 81530003, 81890994, and 81770153) and the National Key Research and Development Program (grant 2019YFA0905900).

Contribution: K.W., Z.C., and Y.T. designed the experiments; Y.T., X.W., H.S., R.Z., Y.Z., and W.J. performed the experiments; Y.T., H.F., X.W., and S.L. analyzed the data; K.W. supervised the study and analysis; K.W., Y.T., and H.F. wrote the manuscript; Z.C. and S.C. gave conceptual advice; Z.C. revised the manuscript; and all authors discussed the results and implications and reviewed the manuscript.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Kankan Wang, Shanghai Institute of Hematology, State Key Laboratory of Medical Genomics, Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, 197 Ruijin Er Rd, Shanghai 200025, China; e-mail: kankanwang@shsmu.edu.cn; Zhu Chen, Shanghai Institute of Hematology, State Key Laboratory of Medical Genomics, Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, 197 Ruijin Er Rd, Shanghai 200025, China; e-mail: zchen@stn.sh.cn; and Hai Fang, Shanghai Institute of Hematology, State Key Laboratory of Medical Genomics, Ruijin Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, 197 Ruijin Er Rd, Shanghai 200025, China; e-mail: hfang@well.ox.ac.uk.