Construction of a single-cell transcriptomic map of stepwise leukemogenesis in a murine AML model

Since MYC is frequently amplified and/or overexpressed in human AML (13%) and in MDS which may develop AML later and was previously called as preleukemia, we constructed an Myc-induced murine AML model (S1A–S1C Fig) [13]. To dissect the whole process of leukemogenesis, we performed a series of scRNA-seq analyses of bone marrow (BM) cells from mice transplanted with leukemogenic HSPCs at different stages (S1D Fig). All the recipient mice developed full-blown AML approximately 8 weeks after transplantation with minimal variation (S1E Fig). The leukemic cells could be identified by the expression of GFP, which was linked with Myc. A majority (80.3%) of the bone marrow in all transplanted mice was leukemia cells that expressed markers of myeloid lineage (B220-; CD3-; Mac-1+; c-Kit+), and over time, the proportion of myeloid cells increased (S1F and S1G Fig). Peripheral blood smears showed step-by-step leukocytosis with increased blasts (S1H Fig). The histological analysis showed that the resulting mouse AML recapitulated the features of human disease (S1I Fig). During leukemogenesis, we harvested the initiating Myc-overexpressing HSPCs (pretransplanted, T₀), preleukemic BM cells at 2 weeks (T₁), 4 weeks (T₂) after transplantation, and full-blown AML BM cells (T₃, 8 weeks after transplantation) for 10× Genomics scRNA-seq analyses, and a total of 41,078 cells were retained for analysis. Two independent repeats for each time point were highly similar to each other (S2A Fig).

Accordingly, we calculated the UMAP plot of all these preleukemic and leukemic cells (PLCs) together with normal BM cells (Fig 1A). The normal cell subpopulations were clustered by their distinct gene signatures, and the remaining GFP-positive cells were defined as the PLC subtype (S2B Fig). While all the normal cell lineages from each time point largely overlapped, PLCs displayed a shift from being close to neutrophils to being close to HSPCs on the UMAP (Fig 1B). The ratios of normal populations gradually decreased, while those of PLCs increased over time (S2C Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Serial scRNA-seq analyses revealing the landscape of progressive leukemogenesis.

(A) UMAP plot of normal cells and PLCs for all 9 samples at 5 time points (normal whole bone marrow (2×), T0 (before transplant, 1×), T1 (2 weeks after transplant, 2×), T2 (4 weeks after transplant, 2×), and T3 (8 weeks after transplant, 2×)) of leukemogenesis. Cells are colored according to Myc-GFP expression (upper left corner). (B) UMAP plot of normal cells and PLCs during leukemogenesis ordered by time point. (C) Density scatter plots showing G2 M scores and S scores in PLCs during leukemogenesis. (D) Density scatter plots showing differentiation and stemness signature scores in PLCs during leukemogenesis. (E) Bar graph showing the number of UMI counts of Myc-GFP-positive and Myc-GFP-negative cells at 3 time points; p values were calculated by the Wilcoxon signed-rank test. (F) The heatmap displaying the normalized expression levels of high-risk MDS, MDS_AML, and AML signatures in PLCs at each time point. The underlying data for Fig 1C, 1D, 1E and 1F can be found in S1 Data. AML, acute myeloid leukemia; MDS, myelodysplastic syndrome; PLC, preleukemic and leukemic cell; scRNA-seq, single-cell RNA sequencing; UMI, unique molecular identifier.

https://doi.org/10.1371/journal.pbio.3002088.g001

Since it has been proposed that uncontrolled proliferation and blocked differentiation are 2 major mechanisms of leukemogenesis [7], we analyzed the cell cycle and differentiation status of PLCs. The results showed a gradual increase in the cell cycle for PLCs from T₀ to T₃ (Fig 1C). PLCs also acquired progressively increased differentiation blocks, as indicated by reduced differentiation scores and enhanced stemness scores from T₀ to T₃ (Fig 1D). We also found that the number of unique molecular identifiers (UMIs) and genes progressively increased in GFP-positive cells, while those of GFP-negative cells did not, which was consistent with a previous report that RNA abundance is associated with elevated expression of stem genes [14] (Figs 1E and S2D and S2E). Overall, we delineated a single-cell landscape of leukemogenesis that recapitulated the stepwise transition of this disease.

Clinically, more than 30% of MDS cases progress to AML. To visualize the molecular switches in PLCs during leukemogenesis, we performed correlation analysis of the transcriptome between PLC populations at different time points and patients with high-risk MDS, MDS-AML, and de novo AML [15]. Of note, the gene signature of T₁ PLCs was highly expressed in patients with high-risk MDS, that of T₂ PLCs was highly expressed in AML patients who developed from MDS and that of T₃ PLCs was highly expressed in de novo AML (Fig 1F). The correlation of the stepwise molecular switches of PLCs with those of patients suggested that the molecular trajectory of PLC transformation in mice represented a potential mechanism of leukemogenesis from MDS to AML patients.

Progressively deteriorated RNA splicing abnormality during leukemogenesis

To dissect the molecular events underlying stepwise leukemogenesis, we performed gene set variation analysis (GSVA) for specific pathways of PLCs of each stage. We found that the Myc target genes gradually increased from T₀ to T₃, along with genes involved in oxidative phosphorylation and DNA repair, and the genes of the p53 pathways gradually decreased over time, all of which were consistent with the gradually increased aggressiveness of PLCs over time (Fig 2A).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The Myc targets, not Myc, triggered a progressively deteriorating RNA splicing abnormality during leukemogenesis.

(A) Heatmap showing GSVA scores of hallmark pathways at each time point during leukemogenesis. (B) The top 10 KEGG pathways of Myc_targets detached in PLCs. (C) The box plot showed the relative expression levels of SFs in PLC during leukemogenesis; p values were calculated by the Wilcoxon signed-rank test. (D) Heatmap showing the relative expression levels of 4 subtypes of MYC targets involved in high SFs (rows) during leukemogenesis (columns) in murine AML. (E) Kaplan–Meier survival curves of TARGET-AML patients with low or high SFs expression at T0, T1, T2, and T3. The p value was calculated by the log-rank test. The underlying data for Fig 2C and 2E can be found in S1 Data. AML, acute myeloid leukemia; GSVA, gene set variation analysis; PLC, preleukemic and leukemic cell; SF, splicing factor.

https://doi.org/10.1371/journal.pbio.3002088.g002

Since myc was the sole driver in this AML model, we first checked the expression levels of Myc. We found that the expression of either ectopic or endogenous Myc was not constitutively increased from T₀ to T₃ (S3A Fig). However, we observed that the expression levels of Myc targets constitutively increased from T₀ to T₃ (S3B Fig). Since the expression level of Myc itself was not progressively changed in PLCs, it was unlikely that the increase in Myc targets was just a selection of cells with high levels of Myc, which further suggested that stepwise leukemogenesis was not a result of simply selecting Myc expression. Importantly, by analyzing the transcriptome of the TCGA AML cohort, we found that the signature of Myc target, but not the expression of Myc itself, was associated with the poor prognosis of AML patients (S3C Fig).

Furthermore, the KEGG analysis of the Myc target genes revealed that the spliceosome pathway was the most enriched among the Myc targets in PLCs, along with the cell cycle pathway and RNA transport pathway [16] (Fig 2B). The expression level of splicing factors gradually increased from T₀ to T₃ (Fig 2C), and stage-specific splicing regulatory factors (SFs) were identified by gene expression (Figs 2D and S3D). We found that the expression of these stage-specific SFs associated with PLCs, but not those associated with normal cells, had prognostic value in human AML patients, and the prognostic value of PLC SFs progressively increased from T₁ to T₃ (Fig 2E). Taken together, these data suggested that aggressively increased expression of splicing factors, independent of the Myc expression level, might underlie the progression from preleukemic to fully transformed leukemia in our stepwise model of leukemogenesis.

Abnormal RNA splicing kinetics distinguish leukemogenic initiating cells from their normal counterparts at the tipping point

Diverging from the normal development route is one of the critical steps at the initiation of tumorigenesis [17]. In the landscape of leukemogenesis, we observed a diverging point where HSPCs would differentiate into GMP cells or be transformed into GMP-like preleukemic cells (Fig 3A). Of note, the progenitor-like and GMP-like AML patients’ signatures are mostly expressing in our GMP-like preleukemic cells [18] (S4A Fig). GFP+ cells contained quite a lot of HSC population at the start T0 that were reduced to almost zero at the end time point T3 (S4B Fig). Consistently, the differentiation route from HSPCs to GMP cells was characterized by reduced expression of leukemia signature genes and increased expression of myeloid/neutrophil lineage genes, while in sharp contrast, the transformation route from HSPCs to GMP-like cells displayed increased expression of leukemia signature genes and decreased expression of differentiation genes (Fig 3B). The kinetics of the expression of GMP-specific and GMP-like-specific genes over pseudotime indicated that there was a fate-determining tipping point for Myc-expressing HSPCs to either the GMP differentiation route or GMP-like leukemogenesis route (Fig 3C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. A tipping point for normal differentiation and malignant transformation during leukemogenesis.

(A) UMAP plot of the tipping point for Myc-GFP-positive HSPCs, GMPs, and GMP-like cells. (B) The UMAP plots showing the expression levels of GMP classical markers, Klf2 and Cd177 (top), and GMP-like markers, Mif and Ncl (bottom). (C) The differentiation trajectory from HSPCs to GMP cells (top) and the transformation trajectory from HSPCs to GMP-like cells (bottom) calculated by Slingshot. (D) The RNA velocity at the tipping point. (E) The box plots showed the expression levels of the AML and LSC signatures in HSPCs, GMPs, and GMP-like cells; p values were calculated by the Wilcoxon signed-rank test. (F) The box plots showed the expression levels of the SF signature in HSPCs, GMPs, and GMP-like cells; p values were calculated by the Wilcoxon signed-rank test. The underlying data for Fig 3C, 3E and 3F can be found in S1 Data. AML, acute myeloid leukemia; HSPC, hematopoietic stem and progenitor cell; LSC, leukemic stem cell; SF, splicing factor.

https://doi.org/10.1371/journal.pbio.3002088.g003

Given the importance of RNA splicing during the whole leukemogenesis process, we analyzed RNA splicing kinetics at the tipping point of leukemogenesis by RNA velocity, which has been suggested to indicate developmental potential and cellular dynamics [19]. Consistent with the differentiation tendency of HSPCs into GMPs, HSPCs displayed relatively stationary RNA velocity, and GMPs had high velocity. There was a strong directional flow from HSPCs to GMP. Strikingly, the RNA velocity of GMP-like cells was more stationary than that of HSPCs, and more importantly, the directional flow was from GMP-like cells to HSPCs, which indicated that GMP-like cells at the tipping point may have more potential than HSPCs (Fig 3D). Importantly, we found that at the onset of leukemogenesis, GMP-like cells did not gain a significantly increased AML or leukemic stem cell (LSC) signature (Fig 3E) compared to HSPCs. However, we observed a significant decrease in the expression of RNA SFs from HSPCs to GMPs but a significant increase from HSPCs to GMP-like cells (Fig 3F). These active splicing factors might give preleukemic cells a selection pool for leukemia-promoting molecules.

NPM1 at the tipping point is involved in leukemogenesis

To identify the key signatures leading to the binary cell fate choice at the tipping point, we analyzed the differential expression of genes on the 2 evolutionary routes and then defined the intersection between the low-expressing gene set 2 in the GMP lineage and the high-expressing gene set 4 in the GMP-like lineage as tipping point signatures (TPS), with a total of 30 genes (S4C and S4D Fig). We characterized the TPS pattern in the GMP lineage and GMP-like lineage and found that the KEGG_RIBOSOME pathway associated with tumorigenesis also conformed to this pattern (S4E and S4F Fig).

Furthermore, CRISPR/Cas9 gene editing was used to verify and screen out genes that affect the differentiation and stemness of Myc-overexpressing HSPCs at the tipping point of malignant transformation. To identify specific genes promoting the differentiation of HSPCs, we ranked the sgRNAs by differences in the mean fluorescence intensity (MFI) of Mac-1 or c-Kit staining and the proportion of Mac-1+ or c-Kit+ cells. Based on all detection indicators, Npm1 and Phgdh were finally selected as key molecules for inducing HSPCs malignant transformation during leukemogenesis (S4G and S4H Fig). In addition, the characteristic molecular interaction network analysis showed that Npm1 was a key node gene in the GMP-like cells, as both the differentially activated module gene and differentially expressed gene (DEG) in the GMP-like cells suggested that Npm1 might be a key player in leukemogenesis (Fig 4A). As expected, Npm1 mRNA was gradually up-regulated in the transformation route and malignant progression, but gradually down-regulated in the differentiation route, consistent with our TPS pattern (Figs 4B and S4I). The abnormal expression of NPM1 has been reported to be involved in human leukemogenesis [20], and consistent with the up-regulation of Npm1 expression in GMP-like cells, NPM1 was also significantly up-regulated in AML patients compared to normal samples (Fig 4C). Therefore, we performed further functional studies of Npm1 to verify its effect on the development of leukemia.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Npm1 acted as a tipping point signature to promote AML.

(A) The molecular interaction networks showing the node genes in GMP-like cells. Red represents DAMGs and DEGs, simultaneously; orange represents DEGs; green represents DAMGs; gray represents connecting Steiner genes. (B) The relative expression of Npm1 along the differentiation (left) and transforming (right) trajectories. (C) The box plot showed the expression levels of NPM1 in TCGA-LAML and normal samples. ***p.adj < 0.001; p values were calculated by Wilcoxon signed-rank test. (D) The Kaplan–Meier survival curve of recipient mice transplanted with Myc-GFP HSPCs harboring sgScramble or sgNpm1. **p < 0.01 (log-rank test). (E) The bar graphs showed the MFI of differentiated cells (left) and stem cells (right) at 2 days, 4 days, and 6 days in sgNpm1 samples compared to scramble samples. *p < 0.05, **p < 0.01, *** p < 0.001; P values were calculated by unpaired parametric t test. (F) Scatter plot showing differentiation and stemness signature scores in sgNpm1 and sgScramble samples. Measured by RNA-seq. (G) The box plot showed the expression levels of the GMP-like signature (left) and GMP signature (right) in sgNpm1 and sgScramble samples. *p < 0.05, **p < 0.01, p values were calculated by Chi-squared test. (H) The KEGG pathways enriched in sgScramble samples compared to sgNpm1 samples (top). KEGG pathways enriched in sgNpm1 samples compared to sgScramble samples (bottom). (I) GSEA showing the negative enrichment of HALLMARK_MYC_TARGETS_V1 and HALLMARK_MRNA_SPLICING in sgNpm1 cells compared to sgScramble cells. (J) GSEA showing the negative and positive enrichments of GMP-like and GMP signatures in shNPM1 K562 cells compared to shREN cells. The underlying data for Fig 4C–4J can be found in S1 Data. AML, acute myeloid leukemia; DAMG, differentially activated module gene; DEG, differentially expressed gene; GSEA, gene set enrichment analysis; HSPC, hematopoietic stem and progenitor cell; MFI, mean fluorescence intensity.

https://doi.org/10.1371/journal.pbio.3002088.g004

To validate the function of Npm1 in leukemogenesis, we disrupted it in Myc-overexpressing HSPCs by CRISPR/Cas9 (S5A–S5C Fig) and found that Npm1 loss significantly extended the latency of Myc-driven AML in mice (Fig 4D). Consistently, Npm1 deficiency promoted the differentiation of HSPCs in vitro, as indicated by significantly increased expression of the differentiation marker Mac-1 and decreased levels of the stem cell marker c-Kit (Fig 4E). Transcriptomic analysis by RNA-seq also showed that Myc-overexpressing HSPCs promoted cell differentiation while decreasing stemness, with more GMP gene signatures and fewer GMP-like signatures after the deletion of Npm1 (Fig 4F and 4G). KEGG also showed that cell differentiation and apoptosis pathways were enriched in Npm1-deficient tumor-initiating cells (Fig 4H). GSEA indicated that the MYC_TARGETS_V1 and MRNA_SPLICING up-regulated genes were significantly negatively enriched in HSPCs with sgNpm1 compared to those with sgScramble, which corroborates our previous analysis of global molecular regulatory events behind the evolution of leukemia from T0 to T3 (Fig 4I). Consistent with our data, the GMP gene signature was significantly positively enriched in NPM1-deficient human leukemic cells compared to control shREN cells, and the GMP-like gene signature was significantly negatively enriched, suggesting that NPM1 might be a driver of leukemia (Fig 4J).

Phgdh, although rarely reported in AML, was another top gene in the tipping point of GMP-like cells. Its expression was up-regulated along the transformation route but down-regulated in the differentiation route (S5D Fig). Furthermore, Phgdh disruption promoted the differentiation of HSPCs (S5E Fig). We also observed that Myc-overexpressing HSPCs reduced the GMP-like gene signature, cell stemness and the cell cycle after Phgdh deletion (S5F and S5G Fig). KEGG also showed that lineage differentiation pathways were significantly enriched in Phgdh-deficient preleukemia cells, and GSEA showed that the MYC_TARGETS_V1 and MRNA_SPLICING up-regulated genes were also significantly negatively enriched in HSPCs with sgPhgdh (S5H and S5I Fig). High expression of PHGDH was associated with poor prognosis in AML patients (S5J Fig). We also have observed up-regulated expression of NPM1 and PHGDH in most malignancies compared to normal samples (S5K and S5L Fig). As outlined above, our data suggested that the AML tipping point signatures, including Npm1 and Phdgh, might be a critical driver for leukemia initiation at the tipping point.

RNA splicing creates explosive heterogeneity of leukemic cells

Once the PLCs went through the tipping point and embarked on the tumorigenesis route, they gradually gained malignancy and heterogeneity, which could be visualized on the URD map (S6A Fig). All these PLCs could be grouped into 5 major subpopulations, which were named after their top marker genes (Fig 5A and 5B). On the map were all the T0 cells at the “root,” followed by T1 and T2 cells, sequentially, at the “stem.” T3 cells were located at the crown and ended with multiple “branches” (Fig 5C). The expression of the human AML signature genes gradually increased along with the leukemogenesis “tree” (S6B Fig). The ratios of these subpopulations significantly varied among the time points, and Hsp90aa1^hi cells did not emerge until T3, suggesting that this subpopulation might be more aggressive than others (S6C Fig). Consistently, distinguished expression modules were identified in PLCs at different stages T0, T1, T2, and T3, suggesting that the molecular features of PLCs are changed during leukemogenesis (S6D Fig). The Hsp90aa1^hi population had the highest expression of stemness score and the Npg^hi population had the highest expression of differentiation score (S6E Fig). Moreover, the 5 clusters represented different differentiation stages: the Hsp90aa1^hi and Ifitm3 ^hi populations were more stemness and the Npg^hi population was well differentiated (S6F Fig). As expected, we found that these subpopulations displayed distinguishable gene signatures in the TCGA_LAML and TARGET_AML databases, and the hazard ratio calculated that the HSP90AA1^hi subtype was significantly associated with high risk in human AML patients and HSP90AA1 expressions were up-regulated in even most malignancies (S6G–S6K Fig). Consistently, we also found a positive correlation between the expression levels of Npm1 or Phgdh gene and the Hsp90aa1^hi subpopulation signatures in tipping point cells (S6L Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Explosive intraleukemia heterogeneity during leukemogenesis.

(A) The force-directed layout maps showing leukemogenesis trajectories, colored by cell subtype. (B) The dot plot showed the expression of cell type-specific genes (column) across cell populations in PLCs. (C) The force-directed layout maps showing leukemogenesis, colored by time series. (D) The box plot showed the expression levels of SFs in each subpopulation; p values were calculated by the Wilcoxon signed-rank test. (E) UMAP plot of 165,33 single cells from bone marrow of 5 AML patients, colored by cell population. (F) The dot plot showed the expression of cell type-specific genes (column) across cell populations in AML patients. (G) Hierarchy clustering showing the orthologous murine (mmu) and human (hsa) PLC subtypes at the single-cell level. (H) The box plot showed the expression levels of cell cycle signatures in PLC subtypes in AML patients; p values were calculated by the Wilcoxon signed-rank test. (I) The box plot showed the expression levels of SFs in PLC subtypes in AML patients; p values were calculated by the Wilcoxon signed-rank test. (J) The t-SNE plot showing the Myc-GFP-positive cells in T3, performed by Smart-seq2, colored by cell types. (K) Kaplan–Meier survival curves of TCGA-LAML patients with low or high Hsp90aa1^hi cell signatures. The p value was calculated by the log-rank test. The underlying data for Fig 5B, 5D, 5F and 5H–5K can be found in S1 Data. AML, acute myeloid leukemia; PLC, preleukemic and leukemic cell; SF, splicing factor.

https://doi.org/10.1371/journal.pbio.3002088.g005

We further explored the potential molecular mechanisms underlying the heterogeneity, especially those promoting aggressiveness in the Hsp90aa1^hi subpopulation, and found that Hsp90aa1^hi cells expressed significantly higher levels of splicing factors than any other subpopulation (Fig 5D), suggesting that the splicing machinery generates leukemia intraheterogeneity and promotes its aggressiveness.

To validate the intraheterogeneity in human AML, we analyzed the single-cell transcriptomes of BM cells from 5 AML patients. The leukemic cells were recognized by referred copy number variations (CNVs) (S7A Fig). These AML cells could be divided into 4 distinct subpopulations (Fig 5E), and the ratios of these subpopulations significantly varied among these 5 patients (S7B Fig). Of note, the HSP90AA1^hi subpopulation shared the same marker genes, HSP90AA1 and PDLIM1, with the Hsp90aa1^hi subpopulation in mouse AML. Similarly, the TYROBP^hi subpopulation expressed the same top marker genes, TYROBP and SAT1, as the Tyrobp^hi subpopulation in mouse AML (Fig 5F). Importantly, through unbiased hierarchy clustering according to their transcriptomes, we found that the human HSP90AA1^hi subpopulation was most similar to the mouse Hsp90aa1^hi subpopulation, the human TYROBP^hi subpopulation was most similar to the mouse Tyrobp^hi subpopulation, and the human MPO^hi and SPINK2^hi subpopulations were close to the mouse Hes^hi subpopulation (Fig 5G). Furthermore, the HSP90AA1^hi subpopulation was the most aggressive among all cells, as indicated by the highest cell cycle score, while the TYROBP^hi subpopulation had the slowest cell cycle in human AML (Fig 5H). Similar to those in mouse AML, the HSP90AA1^hi cells expressed the highest levels of splicing factors, while the TYROBP^hi cells expressed the lowest (Fig 5I). KEGG analysis showed that splicing pathways were also significantly enriched in the HSP90AA1^hi cell population (S7C Fig). The similarity of the intraheterogeneity and its associated molecular features between human and mouse AML suggested that our mouse AML model faithfully recapitulated the molecular and cellular characteristics of human disease and that abnormal RNA splicing generally promoted leukemic heterogeneity in various AML.

Exon 6 skipping of Tmem134, predominantly in the Hsp90aa1^hi subpopulation, promotes aggressiveness in AML

To deeply characterize abnormal RNA splicing events underlying leukemia heterogeneity, we performed single-cell RNA-seq analysis of 178 leukemic cells in T3 by Smart-seq2 instead of the 10× Genomics platform, which would be able to analyze the full length and more transcripts [21,22]. Based on their gene expression patterns, these cells were partitioned into 5 populations, and the t-SNE plot showed that the Hsp90aa1^hi subpopulation was separate from others (Figs 5J and S8A). Indeed, the high expression of the signature genes of the Hsp90aa1^hi cell population was also significantly associated with poor prognosis in human AML patients (Fig 5K).

Consistent with those analyzed by 10× Genomics, the signature genes of Hsp90aa1^hi cells were enriched in the RNA splicing pathway and DNA metabolic process pathway (S8B Fig). Furthermore, we found that all 5 major subtype events of alternative RNA splicing, including exon skipping, intron retention, alternative 5′ splice site, alternative 3′ splice site, and mutually exclusive exons, were increased in Hsp90aa1^hi cells compared with Hsp90aa1^lo cells (S8C Fig). These data validated the RNA splicing in the Hsp90aa1^hi subpopulation, which displayed more aggressiveness.

Among the alternatively spliced genes in Hsp90aa1^hi cells was transmembrane protein 134 (Tmem134), a putative transmembrane protein with 2 transmembrane domains on the C-terminus [23]. The full-length Tmem134, named Tmem134α, and exon 6 skipped Tmem134, named Tmem134β (Fig 6A). TMEM134 is highly conserved in mice and humans, and importantly, the same exon skipping also occurred in human AML (Figs 6B and S8D). Furthermore, we also measured the PSI values of TMEM134 exon 6 skipping in multiple cancers and found that it happens in the most of them more or less, suggesting that it may have the general function in pan-cancer (S8E Fig). We found that exon 6 of Tmem134 was skipped more in Hsp90aa1^hi cells than in other cells (Fig 6C). Importantly, exon 6 skipping of TEME134 was also validated in multiple human AML cell lines and 6 AML patients (Fig 6D and 6E). In addition, HSPCs with TMEM134α overexpression hindered leukemogenesis in our Myc-induced leukemia murine model compared to the empty vector, while TMEM134β significantly promoted leukemogenesis (Figs 6F and S9A). We observed that compared to TMEM134α, TMEM134β-overexpressing tumor-initiating cells proliferated rapidly in vivo with a strong growth advantage (S9B Fig). Similarly, TMEM134β promoted Myc-overexpressing HSPC growth, while TMEM134α inhibited HSPC growth ex vivo (Fig 6G).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. The exon 6 skipping of Tmem134, associated with the most aggressive Hsp90aa1^hi subpopulation, promoted the cell cycle of leukemic cells.

(A) Schematic diagram showing the frequent occurrence of homologous skipping events of TMEM134 exon 6 in murine and human AML. (B) The percentages of TMEM134 exon 6 skipping occurrences in normal and AML patients. Normal (n = 3), AML patients (n = 8). (C) The bar graph showed the Tmem134 exon 6 skipping ratio in Hsp90aa1^lo and Hsp90aa1^hi cells. *p < 0.05 (Wilcoxon signed-rank test). (D) Semiquantitative PCR showing of the relative mRNA level of TMEM134α or TMEM134β in human AML cell lines. (E) Semiquantitative PCR showing of the relative mRNA level of TMEM134α or TMEM134β in 6 human AML patients. Overview of AML patients including gender, age, blast percent was shown above and genetic alterations was listed in S3 Table. (F) The survival curves of recipient mice transplanted with Myc-overexpressing c-Kit+ HSPCs overexpressing vector, TMEM134α or TMEM134β. **p < 0.01 (log-rank test). (G) Growth curves of c-Kit+ HSPCs overexpressing vector, TMEM134α or TMEM134β under Myc at the indicated time points. **p < 0.01, ***p < 0.001. P values were calculated by unpaired parametric t tests. (H) The growth curves of endogenous Tmem134 KO mouse AML cells overexpressing vector, TMEM134α or TMEM134β at the indicated time points. *p < 0.05, **p < 0.01. P values were calculated by unpaired parametric t tests. (I) Growth curves of Cas9⁺ AML cells transduced with sgScramble, sgExon6, or sgIntron6+7 at the indicated time points. **p < 0.01, ***p < 0.001. P values were calculated by unpaired parametric t tests. (J) The growth curves of endogenous TMEM134 KO THP-1 cells overexpressing vector, TMEM134α or TMEM134β at the indicated time points. **p < 0.01, ***p < 0.001, p values were calculated by unpaired parametric t test. (K) The box plot showed the relative expression levels of TMEM134α up- and down-regulated genes compared to the vector in the Hsp90aa1^lo and Hsp90aa1^hi subpopulations. Each bar represents the mean ± SD. ****p < 0.0001, p values were calculated using an unpaired parametric t test. (L) GSEA showing that the TAKDDA_TARGETS_of_NUP98_HOXA9_FUSION_UP gene set was negatively enriched in TMEM134α AML cells compared to vector AML cells. The underlying data for Fig 6B, 6C and 6F–6L can be found in S1 Data. AML, acute myeloid leukemia; GSEA, gene set enrichment analysis; HSPC, hematopoietic stem and progenitor cell.

https://doi.org/10.1371/journal.pbio.3002088.g006

Furthermore, we tested the potential effects of exon 6 skipping of TMEM134 in Myc-induced AML cells ex vivo. To exclude the influence of endogenous TMEM134 protein, we generated a Tmem134 knockout Myc AML monoclonal cell line (i.e., M#8) with CRISPR/Cas9 and then introduced TMEM134α or TMEM134β cDNA into M8# by retrovirus (S9C–S9G Fig). The results showed that without endogenous TMEM134, exogenous TMEM134α significantly reduced the growth of leukemic cells compared to the empty vector, while TMEM134β significantly promoted the growth of leukemic cells (Fig 6H). In addition, we used 2 different strategies by abrogating the splicing site of exon 6 (sgExon6) or deleting the whole exon 6 (sgIntron5 combined with sgIntron6) for enforced exon 6 skipping of Tmem134 in leukemia cells (S9H Fig). Both sgExon6 and sgIntron5+6 significantly increased the growth of AML cells (Fig 6I). Furthermore, we introduced synonymous mutated TMEM134α/β cDNA into the THP-1 and HL-60 cell line (2 kinds of human leukemia cell lines) with endogenous TMEM134 knockout (S9I–S9K Fig). Consistently, TMEM134α significantly repressed the growth of human leukemia cells, and TMEM134β was promoted (Figs 6J and S9L). Then, we tested the in vivo effects of TMEM134 exon 6 skipping on leukemia maintenance by transplanting leukemic cells harboring TMEM134α or TMEM134β into recipient mice. We found that recipients with TMEM134α had significantly extended survival (S9M Fig). Despite the short period of tumor development, we have not yet had time to observe the in vivo growth advantage of TMEM134β-overexpressing leukemia cells after the second transplant. Thus, exon 6 skipping of TMEM134, which likely abrogated the tumor suppression function of full-length TMEM134, drove aggressiveness in AML.

Consistently, EdU incorporation assays showed that the proliferation ratios of cells with sgExon6 or sgIntron5+6 were significantly higher than those of the control cells (S9N Fig). In both human and mouse AML, TMEM134α was significantly associated with a decreased cell cycle (S9O Fig). Transcriptomic analysis by RNA-seq also showed that TMEM134α significantly repressed the expression of cell cycle genes in AML (S9P Fig). The genes up-regulated by TMEM134α were expressed at significantly lower levels in Hsp90aa1^hi cells than in Hsp90aa1^lo cells, and in contrast, the genes down-regulated by TMEM134α were expressed at significantly higher levels in Hsp90aa1^hi cells than in Hsp90aa1^lo cells (Fig 6K). GSEA indicated that the NUP98-HOXA9 up-regulated genes were significantly negatively enriched in leukemia cells with TMEM134α compared to those with empty vector (Figs 6L and S9Q). These data strongly indicated that exon 6 skipping of TMEM134 was critical for generating leukemia intraheterogeneity by promoting the aggressiveness of a subpopulation.

Mice

C57BL/6 mice (Jackson Lab, Cat# 000664) were used as HSPC donor mice and transplantation recipients. CAG-Cas9-EGFP mice (Cat# 026179) were purchased from the Jackson Laboratory. All animal studies were approved by the Institutional Animal Care and Use Committees of Sichuan University (20220301026).

Murine AML model and AML cell line

Whole bone marrow (WBM) cells were obtained from 6-week-old C57/BL6 mouse tibiae, femora, and ilia bones and stained with mCD117 magnetic microbeads (Miltenyi Biotec, Cat# 130-091-224). Then, cKit+ HSPCs were enriched from MidiMACS Separator (Miltenyi Biotec, Cat# 130-042-302) and cultured in BCM medium (50% DMEM + 50% IMDM medium) supplemented with 20% FBS, 0.34% beta-mercaptoethanol, 10 ng/mL mIL3/mIL6, and 50 ng/mL mSCF. After 24 h of cultivation in the incubator, the retrovirus harboring pMIG-Myc was spin infected and centrifuged at 2,000 rpm for 1 h. After 24 h, cKit+ HSPCs with an approximately 20% infection ratio was injected via the tail vein into sublethally irradiated (5.5 Gy) 7-week-old female wild-type C57/BL6 mice (1*10^6 cells per mouse). Bone marrow-transplanted mice were maintained on medicated water (ciprofloxacin, 100 μg/mL) for 2 weeks and monitored for leukemogenesis. The mice were observed for a period of 7 to 8 weeks and were sacrificed by cervical dislocation when they became moribund. Statistical analysis of all survival data was carried out using the log-rank test from Prism 9. AML cells were harvested from the bone marrow or spleen and cultivated into a cell line in BCM medium.

Cas9 AML cell line

pLentiCRISPV2 (Addgene plasmid #52961) was transduced into AML cells by lentivirus generated by the 293T cell line, and then Cas9-expressing AML cells were selected by puromycin administration (1 μg/mL) for 1 week.

Retroviral constructs

In Fig 6F and 6G, the cDNA elements overexpressing TMEM134α and TMEM134β were cloned into pMSCV-IRES-GFP-Myc (pMIG-Myc). In Figs 6H and 6J and S9L and S9M, the cDNA elements overexpressing MYC, TMEM134α, and TMEM134β (synthesized by Sangon Biotec, Shanghai) were cloned into pMSCV-IRES-GFP (pMIG) as previously reported [30] and pMSCV-IRES-mCherry (pMIC, constructed in house) retroviral vectors harboring a GFP/mCherry fluorescent reporter. Retroviruses were generated in the 293T cell line, and cDNA expression constructs were transduced into the cKit+ HSPC/AML cell line by a retrovirus.

CRISPRs

CRISPRs were designed at https://www.atum.bio and then cloned into V2TC (modified by pLentiCRISPV2; briefly, the Cas9 component in the vector was replaced with a mCherry fluorescent reporter). The target sgRNA sequences are shown in S1 Table. Lentiviruses were generated in 293T cell lines, and sgRNAs were transduced into Cas9-expressing cKit+ HSPCs/AML cell lines by lentivirus. The PCR primers are shown in S2 Table.

Cell growth curve

Cells were seeded at a concentration of 1*10^4 or 2*10^4 cells per well of a 96-well plate with 3 wells of each line. Cells were counted every 2 days, and subsequent cell counts were normalized to the seeding density after each count. All curves were obtained at least 2 independent times.

Cell proliferation and cell differentiation

Cell proliferation assays were performed with a Cell-Light EdU Apollo488 In Vitro Kit (RIBOBIO, Cat# C10310-3). Briefly, cells were infected with viruses encoding Tmem134α or Tmem134β cDNA or empty vector. On day 3 after infection, cultured cells were stained with 1:4,000 EdU staining solution for 2 h, and then the cells were washed and fixed in 4% PFA. Cells were then incubated with 0.5% Triton X-100 and stained with Apollo reaction buffer.

For the cell differentiation experiment, Cas9-expressing HSPCs were infected with viruses encoding Myc and sgRNAs targeting scramble or Npm1. On days 2, 4, and 6 of infection, cultured cells were analyzed by staining for the differentiation surface marker Mac-1 (Biolegend, clone: M1/70, Cat# 101223)/Gr-1 (Biolegend, clone: RB6-8C5, Cat# 108430) and the stemness surface marker c-Kit (Biolegend, clone: 2B8, Cat# 105814). Double-positive and single-positive populations were gated following standard flow analysis to detect c-Kit and Mac-1/Gr-1 expression levels.

All flow cytometry analyses were performed with an LSRFortessa instrument (BD Biosciences).

Western blot

Cells were lysed with cell lysis buffer (Cell Signaling Technology, Cat#9803), and cell lysate was clarified by centrifugation, according to the manufacturer’s protocol. Anti-NPM1 (Proteintech, clone: 4F12A3, Cat#60096-1-Ig), Anti-PHGDH (Proteintech, polyclonal, Cat#14719-1-AP), and Anti-TMEM134 (Invitrogen, polyclonal Cat#PA5-60492) were used for western blotting. Actin (Invitrogen, clone: BA3R, Cat#MA5-15739-HRP) was used to monitor equal loading.

RNA-seq analyses

RNA-seq libraries were constructed by an Illumina Stranded mRNA Sample Preparation Kit (NEB, Cat# E7770) according to the manufacturer’s protocol and were sequenced by an Illumina NovaSeq 6000 sequencing machine with 150-bp paired-end reads. The RNA-seq reads were aligned to the mouse reference genome (GRCm38) by STAR_2.6.0a. Transcript abundance was normalized and measured by transcripts per kilobase million (TPM). GSEA was applied to show statistically significant similarities and differences between 2 given clusters by identifying a priori defined gene sets.

SMART-seq2 library preparation and sequencing

SMART-seq2 was performed on AML cells at T₃, following a modified SMART-seq2 protocol published previously [31]. The libraries were sequenced by an Illumina HiSeq X Ten Sequencing machine with 150-bp paired-end reads. The paired-end reads were aligned to the UCSC mm10 references by STAR_2.6.0a.

Single-cell collection, library preparation, and sequencing

WBM were isolated from mouse tibiae, femora, and ilia leg bones by crushing in 2% FBS (PBS) buffer, add 1:1,000 DNase1 and then translate into 15 mL BD tube, centrifuge at 1,200 rpm, RT for 3 min. Then, the supernatant was discarded, 3 mL of 1x ACK lysis buffer was added to resuspend the bone marrow pellet on ice for 3 to 5 min, and PBS (more than lysis buffer) was added to terminate the lysis reaction. The supernatant was discarded, and 0.04% PBS-BSA buffer was added to resuspend the cell pellet, which was then filtered through a 40-μm cell strainer. More than 80% of the living cells in the sample and a density of living cells ranging from 600 to 1,200 cells/μl were preferable. Sample quality was measured using a Bio-Rad TC20 automated cell counter. Libraries were prepared using Chromium Single Cell 3′ Reagent Kits v2 according to the manufacturer’s protocol (10× Genomics) and were sequenced using Illumina NovaSeq 6000 sequencing.

Single-cell gene expression quantification

The ectopic Myc sequences were added to the reference genome, which could distinguish exogenous and endogenous cells in AML leukemogenesis. Cellranger (v2.1.1) was used to align clean reads with mm10. The Seurat (v.3.2.0) pipeline was applied to analyze scRNA data. Genes expressed in less than 5 cells were not considered, and cells that contained lower than 10% mitochondrial genes and contained at least 200 but not more than 6,000 genes were retained for the following analysis. Normalized data and scale data were generated from the Seurat workflow.

Batch effect reduction, dimension reduction, and unsupervised clustering

To integrate all the single cells from different samples for unsupervised clustering, integration anchor algorithms in Seurat were used to reduce the batch effect. The top 4,000 variable genes were used as highly variable genes by FindIntegrationAnchors, and an integrated assay matrix was used to build the nearest neighbor graph.

For dimension reduction, UMAP, t-SNE, diffusion-map, and force-directed graph were calculated on a batch-reducing matrix. UMAP and t-SNE were calculated by Seurat, and diffusion maps and force-directed graphs were calculated by URD. The cell subtypes were identified by the classic markers in previous studies [32,33].

To unify the identities of cell subtypes from a different platform, we used TransferData in Seurat. The 10× genomics data were used as the reference files, and smartseq2 data were used as query files to carry out transfer label pipelines. Then, unsupervised clustering algorithms were used to identify each cell identity in smartseq2.

Trajectory construction and pseudotime calculation

To calculate the trajectory and pseudotime on the UMAP plot of tipping point cells, the position matrix of scRNA data was extracted and processed to feed on Slingshot. Three hundred fixed points were performed to fit an approximate trajectory. In addition, the general additive model was used to identify the genes with dynamic expression based on pseudotime. The force-directed layout and pseudotime of PLCs were determined by URD packages.

GSVA, gene regulatory network analysis, and molecular interaction network analysis

GSVA was used to calculate the enrichment score per cell or sample. ClusterProfiler was used to enrich the GO and KEGG pathways. Pathway network analysis and visualization were performed on the modified cnetplot function. The molecular interaction network was calculated by RSCORE, with mouse-specific protein–protein interaction networks downloaded from BioGRID. Gene interaction matrix, generated from the gene expression matrix by R. SCORE function. The gene interaction network was visualized by the Steiner tree. The genes with highly significant differential expression and genes with highly significantly different network interactions are highlighted in the network map.

Generation of gene signatures and visualizations

Cell type-specific gene signatures were generated by the FindMarker/FindAllMarker function implemented in Seurat. Stemness gene signatures and differentiation gene signatures were generated from HSPC-specific expressed genes or neutrophil-specific expressed genes with significant statistical power. Gene sets of Myc targets and splicing factors were downloaded from the Molecular Signatures Database (MSigDB). The mean value of the log2-transformation matrix was used to quantify the global activation of gene signatures or pathways in each cell/sample. The signature score was projected on the UMAP plot by using the modified FeaturePlot function.

Splicing events calculation

rMATS v4.0.2 was used to analyze splicing events on Smart-seq2 single-cell RNA-seq data. bam files were divided into 2 groups based on the cluster information. We summarized all splicing event types of rMATS results, including SE, A3SS, A5SS, RI, and MXE.

Cell cycle score calculation

To calculate the cell cycle score in each single cell, cell cycle phase signatures for humans and mice were obtained from the Seurat pipeline [34]. In each cell of single-cell data, CellCycleScoring with nbin = 28 was used to calculate G2 M scores and S scores. Cells with positive scores were divided into the G2 M phase or S phase, and their G1 scores were assigned to 0. For cells with negative scores, G1 scores were calculated by the following formula:

After the G1 scores were calculated, the R package ggpubr was used for visualization. To calculate G1 scores in each bulk RNA-seq sample, the z score of the mean TPM value of G2 M signature scores and S phase signatures was evaluated.

Bulk data processing and signature identification in MDS and AML patients

Low-risk, high-risk MDS and MDS_AML patient data GSM2430698 [15] and AML patient data GSE1159 [35] were downloaded from the GEO database. We processed MDS patient data by the RNAs-seq workflow as previously mentioned. DESeq2 was used to identify the DEGs. Three types of signatures were obtained, including low-risk MDS signatures generated by comparing low-risk samples to normal samples (p.adj < 0.05 and log2FoldChange > 3), high-risk MDS signatures (p.adj < 0.05 and log2FoldChange > 4), and MDS-AML signatures (p.adj < 0.05 and log2FoldChange > 4) generated by the same kind of comparison. For AML patient data, DESeq2 was used to identify the DEGs (p.adj < 0.05 and log2FoldChange > 1.5). To project the patients’ signature on mouse data, we used the R package biomaRt to convert the human genes to homologous mouse genes.

TCGA data analysis

Clinical information and transcriptome data of TARGET-AML and TCGA-AML were downloaded by the R package TCGAbiolinks. The percent-splice-in (PSI) value of TCGA-LAML was downloaded from TCGA SpliceSeq. Then, the clinical data were processed by filtering patients who could not match the transcriptome and clinical information. Patients whose follow-up days and days to death were not unknown were filtered.

To test the correlation of selected signature expression and patient survival, we calculated the signature scores in each patient based on the log2 transformation of the FPKM matrix. Then, surv_cutpoint was used to group patients based on the signature scores. Then, the Surv function and survfit function were used to fit reasonable survival curves, and the ggsurvplot function was used for visualization. The statistical power of survival differences was calculated by chi-squared distribution.

To test the correlation between splicing events and patient survival, the PSI value from TCGA SpliceSeq was used to divide patients into 2 groups. In the same way mentioned above, the survival curve and statistical power were visualized and quantified.

Identification of malignant cells with CNV estimation

To distinguish malignant cells from nonmalignant cells in human AML generated by 10× Genomics, we inferred large-scale chromosomal CNVs in each single cell based on moving averaged expression profiles across chromosomal intervals [36].

RNA velocity analysis

Velocyto was used to quantify the unspliced and spliced counts from BAM files in scRNA data by Run10x mode. The mm10 genome and mm10_rmsk references were downloaded from UCSC and used for alignment. RNA velocity analysis was quantified by velocyto. R program.

Ethical approval

All mouse experiments were performed in compliance with the Guide for the Care and Use of Laboratory Animals of Sichuan University and were approved by the Animal Care and Use Committee of Sichuan University.