G-quadruplexes (G4) are four-stranded, noncanonical secondary DNA structures formed in guanine-rich sequences15. G4 was reported to be related to the R-loop formation16,17. Both structures show accumulation in G-rich regions and cause DNA double-strand breaks (DSBs) by inducing genome instability14. DEAD box family proteins, such as DDX5, were shown to be essential to resolve DNA G410,18. G4 homeostasis is also critical for transcriptional regulation during development19. Through in vivo studies and human models, we show that DDX41 resolves G4 structures in the erythroid genome. DDX41 deficiency leads to G4 accumulation, causing genome instability, ribosomal defects, and cGAS-mediated cell death, impairing erythropoiesis and contributing to myeloid neoplasm pathogenesis.
Loss of Ddx41 in mouse hematopoietic cells in vivo was shown to be embryonically lethal with unclear cellular and molecular mechanisms. To investigate Ddx41’s functions in hematopoiesis, we first generated a hematopoietic-specific Ddx41 knockout mouse model by crossing floxed Ddx41 (Ddx41) with VavCre mice (Supplementary Fig. 1A). As reported, Ddx41 deficiency led to embryonic lethality, although VavCre expression in oocytes and endothelial cells raises the possibility that the observed lethality may involve contributions from non-hematopoietic lineages. Morphologic examination at E14.5 showed that VavCre:Ddx41 embryo was severely pale, indicating defects in erythropoiesis (Supplementary Fig. 1B). We next purified Ter119 (a mature red cell marker) negative fetal liver HSPCs from the mutant fetuses and cultured the cells in an erythropoietin (Epo)-containing erythroid differentiation system. We found that loss of Ddx41 completely blocked erythroid differentiation with a marked increase in cell death (Supplementary Fig. 1C and 1D), demonstrating that Ddx41 is essential for erythropoiesis.
To study the role of Ddx41 in erythropoiesis, we first determined the expression levels of Ddx41 during different stages of erythropoiesis in an in vitro mouse bone marrow erythroid culture system. We found that Ddx41 is highly upregulated in the early stages of terminal erythropoiesis (day 1 in culture) when the cells rapidly underwent differentiation and proliferation. Its level was then reduced on day 2 when the cells were prepared for enucleation (Fig. 1A). To study Ddx41 in vivo, we generated two erythroid-specific Ddx41 knockout mouse models, EpoRCre:Ddx41 and HBBCre:Ddx41 mice, to test the functions of Ddx41 at different stages of erythropoiesis. EpoRCre:Ddx41 mice manifest Ddx41 deletion at the progenitor stages of erythropoiesis (approximately the CFU-E (colony forming unit-erythroid) stage, day 0 in culture system), whereas HBBCre:Ddx41 mice exhibit Ddx41 deletion at the terminal stages of erythroid differentiation (Fig. 1B and Supplementary Fig. 1E). We found that EpoRCre:Ddx41 mice were also embryonically lethal (Fig. 1C). The lack of appropriate morphogenesis suggests that Ddx41 could be critical for primitive erythropoiesis (Fig. 1D). We next analyzed EpoRCre:Ddx41 mice that survived to adulthood. These mice exhibited normal fetal liver erythropoiesis (Supplementary Fig. 2A). They showed mild macrocytic anemia with normal white blood cell and platelet counts at 2 months old (Fig. 1E and Supplementary Fig. 2B), which partially mimics patients with DDX41 germline mutations before the development of overt MNs upon somatic second hit. We sacrificed the EpoRCre:Ddx41 mice and analyzed bone marrow erythroid cells using flow cytometry of cell surface CD71 and Ter119, two established markers for erythroid maturation. We found significantly compromised terminal erythropoiesis and increased myeloid cells percentage-wise and in absolute numbers with Ddx41 heterozygosity (Fig. 1F and Supplementary Fig. 2C). This was accompanied by increased Ter119 positive erythroid cells in the spleen (Fig. 1G and Supplementary Fig. 2D), which indicates compensatory extramedullary stress erythropoiesis.
In contrast, HBBCre:Ddx41 mice were viable with no detectable hematologic phenotypes, manifested by normal complete blood count (CBC) at 2 months old (Fig. 1H and Supplementary Fig. 2E). To confirm that Ddx41 is knocked out at the late-stage terminal erythroblasts in HBBCre:Ddx41 mice, we cultured HBBCre:Ddx41 bone marrow lineage negative cells in the Epo-containing medium in which mouse erythroid progenitor cells can rapidly differentiate and proliferate to mature red blood cells in 2 days (Supplementary Fig. 2F). Western blotting demonstrated that Ddx41 showed a slight decrease on day 1 when the cells were predominantly at the early stages of terminal erythropoiesis proerythroblast stages. Ddx41 was markedly decreased on day 2 when the cells were at the late-stage terminal erythropoiesis (Fig. 1I). Consistent with the CBC data, flow cytometry analyses showed no difference in the differentiation and enucleation of the cultured bone marrow erythroid cells from HBBCre:Ddx41 and their littermate control mice (Fig. 1J). Similarly, we found no defects in erythropoiesis or the differentiation of other lineages in vivo in the bone marrow and spleen of HBBCre:Ddx41 mice (Supplementary Fig. 2G-2K). Together, these results demonstrate that Ddx41 is critical at the early stages of erythropoiesis but dispensable at the late stages in mice.
We generated additional lineage-specific Ddx41 knockout mouse lines to study its roles in hematopoiesis. These include CD11c-Cre:Ddx41, LysMCre:Ddx41 and MRP8Cre:Ddx41, which knockout Ddx41 predominantly in dendritic cells, monocytes, and myeloid/granulocytes, respectively (Supplementary Table 1). Interestingly, these mice were all viable with no obvious abnormalities in their CBC at young ages. Flow cytometry studies revealed no obvious differences in the hematopoietic tissues in these animals (Supplementary Figs. 3-5), demonstrating that Ddx41 is dispensable in these lineages, particularly at their differentiated stages.
Given the significant role of Ddx41 in erythropoiesis in animal models, we wonder whether the same also occurs in MN patients with DDX41 mutations. We identified an MDS patient carrying a somatic DDX41 G1673 > A mutation (G530D) on the helicase domain, which was indicated to potentially disrupt the ATP binding to DDX41. In addition to the DDX41 mutation, this patient also carries a PRPF8 T4916 > A somatic mutation. The patient has a long-standing MDS with anemia, neutropenia, and occasional mild thrombocytopenia. We purified the patient’s bone marrow CD34+ HSPCs and CD71+ erythroid progenitor cells. Sanger sequencing showed that DDX41 mutation exists in CD34+ cells but at a noise background level in CD71+ cells. In contrast, the PRPF8 mutation is present at an equal allele frequency in both CD34+ and CD71+ cells (Fig. 1K). These findings suggest that DDX41 mutation may impair erythroid differentiation or survival, resulting in the selective depletion of DDX41-mutant clones in the erythroid lineage.
DDX41’s role in erythropoiesis is underscored by studies in zebrafish, where its deficiency leads to anemia through ineffective erythropoiesis, involving DNA damage response pathways. DDX41 is also critical in regulating DNA secondary structures such as R-loops, which frequently co-exist with G-quadruplexes (G4). G4 is known to stabilize R-Loop and regulate transcription by interacting with various chromatin-binding proteins, including several DDX family proteins. The role of G4 in erythropoiesis is unknown.
To study the role of G4 in erythropoiesis and whether the loss of DDX41 affects G4 formation, we first determined G4 levels in the bone marrow erythroid cells in vivo. We found that G4 is significantly increased in Ter119 positive erythroid cells compared to Ter119 negative cells (Fig. 2A). We next cultured mouse bone marrow lineage-negative cells in the Epo-containing medium and tested G4 levels at different differentiation stages. We found that G4 significantly increased on day 1, corresponding to the highest proliferation and replication stage (Fig. 2B, and Supplementary Fig. 6A). The level of G4 decreased on day 2 when over 30% of the cells were enucleated. The same trend of changes in G4 levels was observed in cultured human CD34+ cells toward erythroid differentiation (Fig. 2C, and Supplementary Fig. 2F). Importantly, G4 levels in vivo in the erythroid cells are particularly higher than those in the other hematopoietic lineages (Figs. 2D, E), further indicating a critical role of G4 during erythropoiesis.
Since VavCre:Ddx41 mice are embryonically lethal, we used EpoRCre:Ddx41 and HBBCre:Ddx41 mice to investigate Ddx41’s role in G4 accumulation in vivo. We first analyzed G4 levels in the bone marrow erythroid cells in EpoRCre:Ddx41 mice using a Ter119 and CD44-based gating strategy in which different stages of erythroid precursors can be separated based on CD44 expression (Fig. 2F). We found a mild but statistically significant increase in G4 level in the proerythroblast population (population I) in EpoRCre:Ddx41 mice compared to their wild-type counterparts (Fig. 2G). G4 levels in the late-stage orthochromatic erythroblasts (population V) showed no difference (Fig. 2H). In HBBCre:Ddx41 mice, we did not find significant differences among different populations in the same assay, which is consistent with the lack of phenotypes in these mice. It is also difficult to pinpoint what developmental stage(s) G4 starts accumulating in vivo in these mice since most of the erythroid cells in the marrow are at the late stages of terminal erythropoiesis when G4 level is reduced. Therefore, we purified lineage-negative HSPCs from these mice and cultured them in vitro in Epo-medium. As Ddx41 started to reduce on day 1 (Fig. 1H), the G4 level started to increase. The G4 level returned to the normal range on day 2, as expected (Fig. 2I, and Supplementary Fig. 6B). These results are consistent with the phenotypes of these mice and the adverse effect of increased G4 in early, but not late, stages of erythropoiesis. To further demonstrate the role of DDX41 in reducing G4 levels in human erythroid cells, we knocked out DDX41 through CRSPR/Cas9 in CD34+ human HSPCs using two different sgRNAs. Indeed, this led to a significant reduction in DDX41 protein levels (Fig. 2J) and a marked increase in G4 levels and cell death (Figs. 2K, L). Consistent with these model systems, we also observed increased G4 levels in the bone marrow mononuclear cells of the MN patient with DDX41 mutation (Fig. 2M).
We next applied pyridostatin (PDS), a selective G4 binding small molecule that stabilizes G4 and perturbs G4 homeostasis, in our in vitro erythroid differentiation system. PDS significantly increased G4 levels in cultured day 1 erythroid cells (Fig. 3A), induced dosage-dependent inhibition of erythroid differentiation, and increased cell death (Fig. 3B). Treatment of wild-type mice with a three-day high dose of PDS also significantly compromised bone marrow erythropoiesis in vivo (Fig. 3C), although no anemia was observed due to the long half-life of red blood cells. We next treated wild-type mice with PDS for 6 weeks through subcutaneously implanted pumps that chronically release the compound. Indeed, this treatment led to significant anemia (Fig. 3D). The white blood cell count was also reduced, primarily due to PDS-induced lymphopenia (Fig. 3D, E). As expected, chronic PDS treatment also led to ineffective terminal erythropoiesis and proportionally increased myeloid cells in the bone marrow (Fig. 3F, G).
We performed murine in vitro colony-forming unit (CFU) assays under different concentrations of PDS. We found that erythroid-related colonies, such as BFU-E and CFU-GEMM, decreased more significantly in number than myeloid cell colonies (CFU-GM) (Fig. 3H). The same results were obtained when we performed CFU assays in human CD34+ cells (Fig. 3I). We also knocked out DDX41 using CRISPR/Cas9 in CD34+ cells. We found significantly reduced BFU-E and CFU-E colonies (Fig. 3J), indicating DDX41 could also be involved in stem cell commitment to the erythroid lineage. These results show that the erythroid lineage is more sensitive to G4 stresses, reinforcing our finding that DDX41 deficiency significantly impairs erythropoiesis.
The genome distribution of G4 has been studied in different cell types and various species. The genome-wide G4 localization in hematopoietic cells and whether DDX41 co-localizes with G4 are unclear. To understand this, we performed CUT&RUN assays in purified maturing mouse bone marrow Ter119+ erythroblasts and bone marrow lineage-negative HSPCs. We found a significant co-distribution of G4 and Ddx41 at the genome level both in HSPCs and erythroblasts. The total number of Ddx41 and G4 peaks increased markedly as cells differentiated into erythroid cells, even though the proportion of overlapping peaks remained relatively stable (Fig. 4A, B). This is consistent with the critical role of Ddx41 in erythropoiesis. Their distributions are mainly in the intergenic and intron regions, consistent with a previous report. Notably, as cells differentiated from HSPCs to the erythroid lineage, both G4 and Ddx41 binding showed increased enrichment in the promoter and coding regions (Fig. 4A). Motif enrichment analysis revealed that Ddx41- and G4-bound regions are significantly enriched for binding motifs of key hematopoietic transcription factors, including Runx1, Gfl1b, Gata3, and Myc, in both HSPCs and erythroid cells. (Supplementary Fig. 7A). We further confirmed these genomic studies using a confocal immunofluorescence assay of G4 and Ddx41 in cultured mouse erythroblasts, which indeed showed their partial co-distribution (Fig. 4C). Computational predictions indicate that ribosomal DNAs (rDNAs) are highly enriched for putative quadruplex formation due to their repetitive DNA sequences. Consistent with these reports, we found enrichment of G4 and Ddx41 on rDNAs in both HSPCs and erythroid cells (Supplementary Fig. 7B, C).
The co-occupation of DDX41 and G4 in the erythroid genome and the upregulation of G4 after DDX41 depletion indicate that DDX41 functionally maintains the level of G4. DEAD box family proteins, such as DDX5, were known to resolve DNA G4. To determine whether DDX41 resolves G4, we first performed a pull-down assay using biotin-conjugated G4 oligos. The canonical G4 motif is Gm-Xn-Gm-Xo-Gm-Xp-Gm where each G-tract (Gm, m = 2-4) is separated by loops (Xn, Xo, and Xp), and n, o, and p are the combination of nucleotides of various lengths (up to 7). We designed three different G4 oligonucleotides. Three non-G4 oligonucleotides were used as controls. These oligonucleotides were annealed with biotinylated counterparts and then captured on streptavidin magnetic beads. They were subsequently incubated with lysates from cultured mouse erythroblasts in the presence of K cations. Ddx41 exhibited specific binding to all three G4 structures of distinct topologies (Fig. 4D), indicating its ability to recognize a wide range of G4 structures.
To determine whether DDX41 directly resolves G4, we performed an in vitro fluorescence resonance energy transfer (FRET) assay in which a fragment of G4 DNA is flanked by 6-fluorescein (6-FAM) on the 3′-end and black hole-1 quencher on the 5′-end (Fig. 4E). Recombinant human DDX41 was added in vitro, together with G4-FRET and K cations. Indeed, we found a dose-dependent increase in the 6-FAM fluorescence with the increasing amount of DDX41, demonstrating that G4 was resolved by DDX41 (Fig. 4F, and Supplementary Fig. 8A). The kinetics of DDX41 in resolving G4 was also rapid and dose-dependent (Fig. 4G, and Supplementary Fig. 8B). Notably, in these experiments, the oligonucleotide was pre-folded into a G4 structure under high-K⁺ conditions and then incubated with recombinant human DDX41 in a low-K⁺ unfolding buffer, which limits the spontaneous refolding of G4. As a control, we included oligonucleotides processed identically but without the initial folding step. These unfolded oligos showed persistent fluorescence that remained stable under the assay conditions, confirming that refolding does not occur spontaneously during the assay (Supplementary Fig. 8C). We next tested two of the most common somatic mutations of DDX41, R525H, and G530D, on their influences on G4 binding capacities. We found that R525H mildly compromised the binding, whereas G530D markedly reduced it (Fig. 4H). We further discovered that both mutants lost G4 resolving activity in all three G4s we tested (Fig. 4I). These data establish DDX41 as a G4 resolvase, which is compromised by its loss of function mutations.
DDX41 was reported to bind to various nucleic acid structures, including dsDNA, R-loop, and DNA/RNA hybrids. To determine the relative affinity of DDX41 to these molecules, we performed a fluorescence polarization assay using recombinant DDX41. We found that G4 DNA represents a preferred or high-affinity substrate for DDX41 compared to other molecules (Fig. 4J), indicating that DDX41 may have evolved a specialized role in resolving G-quadruplexes.
Accumulation of G4 is associated with genome instability and defects in ribosome biogenesis due to G4 enrichment on rDNAs. Compromised ribosome biogenesis during erythropoiesis is well known to trigger p53-mediated cell death, which is believed to be the pathogenesis of Diamond-Blackfan anemia (DBA) and contribute to the development of del(5q) MDS. We found increased γ-H2AX (a marker for genome instability) when the cultured mouse bone marrow HSPCs were treated with PDS (Fig. 5A-C). As expected, the transcription of ribosomal RNAs was significantly reduced (Fig. 5D). Defects in ribosomal RNA biogenesis are known to negatively influence the expression of ribosomal proteins. Consistently, we found PDS treatment significantly reduced many ribosomal proteins, including Rps19 (mutated in 25% of DBA patients), Rps14 (haploinsufficiency in del(5q) MDS patients), and Rpl26 (mutated in certain DBA patients). The level of p53 was also increased (Fig. 5E). To directly investigate how DDX41 deficiency affects ribosomal biogenesis in vivo, we used HBBCre:Ddx41 mice since these mice survive and Ddx41 is depleted in the late-stage erythroblasts when the cells are most abundant. We cultured the bone marrow lineage negative HSPCs from these mice in Epo-medium and found significantly decreased transcription of rRNAs on day 1 when Ddx41 starts to reduce, and G4 is significantly increased (Fig. 5F, Fig. 2I). The ribosomal protein levels remain steady on day 1, possibly due to the stability of proteins compared to RNAs, but significantly reduced on day 2 (Fig. 5G). Interestingly, we found no increase in γ-H2AX when Ter119-negative erythroblasts from HBBCre:Ddx41 or EpoRCre:Ddx41 mice were cultured in vitro (Supplementary Fig. 8D). These results indicate that these lineage-negative cells may be adapted to the Ddx41 deficiency during development.
Consistent with these findings in mouse erythroid cells, DDX41 knockout in CD34+ HSPCs led to a similar reduction of various ribosomal proteins and upregulation of p53. The p53 downstream target p21 was also increased (Fig. 5H). We next tested whether the p53 level increased in the patient with DDX41 mutation. We used bone marrow biopsies from the same patient with the G530D mutation and performed an immunohistochemical stain for p53. We found that the p53 level (Fig. 5I) and γ-H2AX (Fig. 5J) significantly increased in the patient with DDX41 mutation compared to the normal control, consistent with the findings in the mouse models.
While p53 mediates many pathologies of red cell-related diseases, it has been documented that overexpression of p53 does not induce overt abnormalities in a transgenic model. To determine whether p53 mediates the major phenotypes of Ddx41 hematopoietic specific deficiency mice as it does in DBA and del(5q) MDS, we took a genetic approach and crossed p53 knockout mice with VavCre:Ddx41 mice. If p53 is critical, p53 deficiency would rescue the lethality of VavCre:Ddx41 mice. However, no surviving VavCre:Ddx41 Trp53 mice were born (Fig. 6A, B). Dissection of the pregnant mice revealed that VavCre:Ddx41 Trp53 embryos remained pale with an underdeveloped fetal liver (Fig. 6A). These results demonstrate that p53 is not essential to mediate ineffective erythropoiesis and cell death in the Ddx41 hematopoietic-specific knockout mouse model.
DDX41 was reported to sense intracellular dsDNA and activate STING to mediate type I interferon response in dendritic cells. Indeed, we found the level of cGAMP, intracellular second messenger in response to cGAS activation, was significantly increased in PDS-treated erythroid cells (Fig. 6C). The downstream targets of the cGAS-STING pathway, including interferon beta (IFN-β) and the NF-κB signaling, were also activated (Fig. 6D, E). We previously revealed that erythroid cells generate transient nuclear openings in the early stages of terminal erythropoiesis, which could further activate cGAS. To test this, we treated the cultured mouse bone marrow erythroid cells with a caspase inhibitor, which blocks nuclear opening. This led to a significant reduction of cGAMP (Fig. 6F), suggesting that nuclear openings contribute to the vulnerability of the erythroid cells to the genome instability induced by G4 upregulation.
We then took a similar genetic approach and crossed cGas knockout mice with VavCre:Ddx41 mice. Intriguingly, the cGas deficiency completely rescued the embryonic lethality of the VavCre:Ddx41 mice. VavCre:Ddx41cGas-/- (DKO) mice showed no evidence of anemia or other cytopenias (Fig. 6G, and Supplementary Fig. 9A, B). The bone marrow hematopoiesis was also intact (Supplementary Fig. 9C). Western blotting of the bone marrow Ter119+ erythroid cells confirmed cGas and Ddx41 deletions but also showed loss of ribosomal proteins and upregulation of p53 in the DKO mice (Fig. 6H), which indicates that G4 upregulation-mediated genome instability leads to parallel activation of the cGas and p53 pathways. Consistent with this indication, the level of phospho-p65 was unchanged (Fig. 6H), whereas G4 and γ-H2AX remained upregulated in the bone marrow Ter119+ erythroid cells of the DKO mice compared to the cells in their littermate control mice (Fig. 6I). The critical role of cGas in mediating Ddx41 deficiency-induced pathogenesis was further evidenced by the resistance of primary erythroid cells from cGas knockout mice to PDS-mediated cell death (Fig. 6J). We next performed a single-cell RNA sequencing assay of the total bone marrow cells from DKO mice and their littermate control (Supplementary Fig. 9D). While non-erythroid hematopoietic populations showed no apparent distinctions, late erythroid cells in DKO mice exhibited altered gene expression, including down-regulation of genes involved in cell death regulation and erythroid hemostasis (Fig. 6K, L and Supplementary Fig. 9E, Supplementary Data 1). Since apoptotic signals are known to be involved in chromatin condensation in terminal erythropoiesis, it is possible that these pathways are altered in the DKO cells.
cGAS was shown to translocate to the nucleus under DNA double-strand break to suppress DNA repair independent of STING. We found that cGas, as well as Sting, was predominantly located in the cytoplasm in the erythroid cells upon PDS treatment, demonstrating a cGas-Sting-dependent pathway and consistent with the upregulation of IFN-β and the activation of NF-κB signaling (Supplementary Fig. 9F). In line with the role of the cGAS-STING pathway, treatment of DDX41 deficient CD34+ cells with a cGAS inhibitor (RU521), a STING inhibitor (H151), or an NF-κB inhibitor (BOT64) significantly rescued cell death (Fig. 6M). Similar to PDS, treatment of the erythroid cells with IFN-β significantly compromised cell differentiation, proliferation, and induced cell death (Fig. 6N).
These data point towards cGAS-STING activation-induced senescence, instead of p53-induced cell death, in mediating ineffective erythropoiesis in DDX41 deficiency. Consistently, we observed increased senescence-associated β-galactosidase activities in DDX41 CRISPR knockout or PDS-treated CD34+ cells in Epo-containing medium, as well as bone marrow mononuclear cells from the DDX41 mutated patient (Fig. 6O).
To extend these findings to the human bone marrow in vivo setting, we applied an induced pluripotent stem cell (iPSC)-derived human bone marrow organoid system resembling primary human bone marrow biopsy samples (Fig. 7A). Whole-mount 3D imaging revealed an endothelial network accompanied by stromal cells along the capillary wall and hematopoietic cells arranged in clusters within and outside the vessels (Fig. 7B). Erythropoiesis in the organoid mirrors primary human bone marrow samples, forming tight erythroid islands (Fig. 7C). A flow cytometry assay further confirmed multilineage hematopoiesis in the organoids (Fig. 7D). To determine whether these organoids can be applied to study human ex vivo bone marrow engraftment, we incubated the organoids with CellVue-labeled donor bone marrow CD34+ cells for three days. The donor cells could be readily detected and surrounded by the recipient hematopoietic cells in the organoids (Fig. 7E). These donor cells could also differentiate into mature hematopoietic cells detected by immunofluorescence and flow cytometry assays (Fig. 7E, F). With this system, we depleted DDX41 through CRISPR in CD34+ cells and incubated these cells with the bone marrow organoids. As expected, the CD71+ erythroid cells, but not CD11b+ myeloid cells, derived from the donor CD34+ cells transduced with DDX41 sgRNA were significantly reduced compared to the control donor cells. Interestingly, the erythroid cells derived from the recipient bone marrow organoids were also significantly reduced, indicating a potential indirect impact from the inflammatory environment due to the cGAS-STING activation (Fig. 7G).
We also treated the organoid with PDS for 24 hours and analyzed erythropoiesis using flow cytometry. Since the organoid already contains erythroid cells at various stages of development, this approach allows us to study the responses of cells at different stages of erythropoiesis to G4 accumulation. Consistent with the results from other experiments, we found a significant decrease in the population of CD34 + CD71+ cells, which include mainly early-stage erythroid cells. However, no significant differences were found in the population of CD34-, CD71 + , and CD235a+ cells, which represent late-stage erythroid cells (Supplementary Fig. 9G).
