Single-Cell RNA-Seq Reveals AML Hierarchies Relevant to Disease Progression and Immunity

Abstract

Acute myeloid leukemia (AML) is a heterogeneous disease that resides within a complex microenvironment, complicating efforts to understand how different cell types contribute to disease progression. We combined single-cell RNA sequencing and genotyping to profile 38,410 cells from 40 bone marrow aspirates, including 16 AML patients and five healthy donors. We then applied a machine learning classifier to distinguish a spectrum of malignant cell types whose abundances varied between patients and between subclones in the same tumor. Cell type compositions correlated with prototypic genetic lesions, including an association of FLT3-ITD with abundant progenitor-like cells. Primitive AML cells exhibited dysregulated transcriptional programs with co-expression of stemness and myeloid priming genes and had prognostic significance. Differentiated monocyte-like AML cells expressed diverse immunomodulatory genes and suppressed T cell activity in vitro. In conclusion, we provide single-cell technologies and an atlas of AML cell states, regulators, and markers with implications for precision medicine and immune therapies. VIDEO ABSTRACT.

Keywords: acute myeloid leukemia; cancer genetics; genotyping; immunity; leukemia stem cells; single-cell RNA-sequencing.

Figure 1.. Identification of cell populations in healthy BM samples

A. BackSPIN clustering of scRNA-seq data for 6,915 hematopoietic cells from normal BM identified 31 clusters of cells with similar transcriptional states. Heatmap shows the pairwise correlation between the average expression profiles of these clusters (rows and columns). Clusters were merged into 15 cell populations based on marker gene expression (right). B. Heatmap shows the expression of 55 selected cell type-specific genes (rows) across 6,195 single cells ordered by the BackSPIN-defined clusters (columns). C. Stacked barplots show the frequencies of BackSPIN-defined cell types in five normal BMs. D. KNN visualization of 6,915 single-cell transcriptomes (points), with similar cells positioned closer together. Points are color-coded by cell type annotations as in C. See also Figure S1.

Figure 2.. Single-cell profiling of AML tumor ecosystems

A. Overview of AML patients and BM aspirate collections. Cell numbers reflect single-cell transcriptomes that passed quality thresholds. For each patient, pie charts indicate time of sample collections, relative to diagnosis, and clinical blast count. B. Chart shows genetic alterations (red) detected in our cohort by targeted DNA sequencing and cytogenetics. C-D. t-SNE plots show single cells from AML556 (C) or AML707B (D) at successive collections. Each plot shows cells from the indicated time point (red) and other time points (gray). t-SNE plots and corresponding H&E stains depict marrows dominated by AML cells at presentation (Day 0), hypocellular marrows with T-cells after chemotherapy (Day 15–18), or repopulating hematopoiesis (Day 31–41). Scale bar 50 μm. See also Table S1.

Figure 3.. Single-cell genotyping by short-read and nanopore sequencing

A. Illustration depicts procedures for acquiring transcriptional and genotypic information from single cells. Nanowell plates and beads generate WTA product wherein each cDNA is appended to a UMI, a cell-specific barcode (CB), and a priming site (SMART). Product is split and used as input for Tn5-mediated scRNA-seq library generation (left) and single-cell genotyping (right). The single-cell genotyping reaction utilizes biotinylated primers to amplify mutational sites in target genes along with corresponding UMI and CB for sequencing. B. Bubble plot depicts the frequency of mutation detection by single-cell genotyping with short-read Illumina sequencing. Detection is more efficient for mutations in highly expressed genes and near the 3’ polyA signal. C. Scatter plot compares mutation frequencies from DNA sequencing (y-axis) or single-cell genotyping (x-axis). Each point corresponds to a mutational site in a specific AML sample. Six examples are highlighted. D. Genome plot illustrates long nanopore reads for three selected TP53 transcripts from AML328. For each transcript, 100 reads are shown (reads were matched by CB and UMI, indicating they came from the same transcript). Black arrow indicates the location of the primer used for amplification. Base mismatches encoding Q144P (orange) or P152R (blue) mutations are indicated. E. Stacked barplot shows the number of cells in which wild-type or mutant TP53 transcripts of indicated lengths were detected by short-read (gray) or nanopore sequencing (green), or both (red). Fragment length was determined from the nanopore data. F. Genome plot illustrates nanopore reads for two selected FLT3 transcripts from AML328. For each transcript, 100 reads corresponding to the ITD or wild-type allele are shown. Black arrow indicates the location of the primer used for amplification. Base insertions representing a newly detected 60 bp ITD are indicated in exon 14 (pink). G. Genome plot illustrates nanopore reads of a fusion transcript from AML707B aligning to the RUNX1T1 (left) and RUNX1 (right) loci. One hundred reads are shown. Nanopore sequencing enabled detection of the fusion without prior knowledge of the junction. H-I. t-SNE plots for AML556 (H) and AML707B (I) as in Figure 2C–D show cells for which wild-type (blue) or mutant (red) transcripts were detected by single-cell genotyping with Illumina sequencing. J. t-SNE plot for AML707B shows cells for which RUNX1-RUNX1T1 fusion transcripts (green) were detected by nanopore sequencing. See also Figure S2 and Table S2.

Figure 4.. Machine learning classifier distinguishes cell types in the AML ecosystem

A. Schematic of machine learning classifiers used to predict AML cell types (Classifier 1) and to distinguish malignant from normal cells in AML tumors (Classifier 2). B. KNN visualization shows single-cell transcriptomes from normal BM (gray; as in Figure 1D). Cells from AML samples for which wild-type or mutant transcripts were detected were projected onto this graph according to their similarity to the normal cells. Boxes depicting the density of projected cells are colored according to the ratio between wild-type and mutant transcripts. Cells with mutant transcripts (red) project along the HSC to myeloid differentiation axis. C. Barplot shows classification of AML cells with mutant transcripts by the first Random forest classifier. The majority are classified as one of six cell types along the HSC to myeloid axis, thereby defining six malignant cell types. D-E. t-SNE plots of AML556 (D) and AML707B (E) as in Figure 2C–D with cells colored by their classification as malignant (red) or normal (grey). F. Scatter plot compares clinical blast counts (y-axis) to the fraction of cells classified as malignant by the machine learning classifier (x-axis). Each point corresponds to a specific AML BM aspirate (n = 27). G. Heatmaps show cell type prediction scores (rows) for all malignant cells (columns) from five representative tumors. Cells in which wild-type and/or mutant transcripts were detected, or that express cell cycle signature genes are indicated below. H. KNN visualizations show single-cell transcriptomes of normal BM cells (gray; as in Figure 1D). Malignant cells from the respective AMLs were projected onto this graph according to their similarity to the normal cells. The density of projected cells (red) conveys the distinct cell type compositions of these tumors. I. Flow cytometry plots show expression of myeloid differentiation markers by the AML samples. See also Figure S3–5.

Figure 5.. AML cellular hierarchies correlate with underlying genetic alterations

A. Genome plot illustrates nanopore reads for four selected FLT3 transcripts from AML419A. For each transcript, 100 reads are shown. Black arrow indicates the location of the primer used for amplification (exon 11). Base mismatches encoding A680V (exon 16; green) or N841K (exon 20; red) mutations are indicated. Base insertions representing a 24 bp ITD are indicated in exon 14 (pink). The mutations do not co-occur on the same transcripts. B-C. Diagrams show AML419 evolution inferred from co-occurrence of mutations in single cells (B) and VAFs from bulk DNA sequencing (C). The most likely model yields one subclone with an A680V mutation, a second subclone with an ITD, and a third subclone that exclusively harbors an N841K mutation. D. Diagram shows FLT3 protein domains and location of mutations. E. Heatmap shows expression of 180 signature genes for the six malignant cell types (rows) in 40 single cells from AML419A (columns). Cells were assigned to subclone A or B, or subclone C on the basis of FLT3 genotypes. F. Heatmap shows expression of 180 signature genes for the six malignant cell types (rows) in 179 AMLs profiled by bulk RNA-seq (columns). Unsupervised clustering revealed seven subsets of patients with different inferred cell type abundances (clusters A-G). G. Charts indicate chromosomal aberrations (top), mutations (middle) and FAB classifications (bottom) for AMLs in F. Correspondence between cell type compositions and genetics is evident. P-values indicate non-random distribution of events between clusters (Fisher’s exact test). n.s., not significant. H. Flow cytometry histograms show expression of the primitive cell marker CD34 in MUTZ-3 cells, four days after transduction with FLT3-WT, FLT3-D835Y, FLT3-ITD or a control gene (luciferase). I. Plot shows change in the percent of CD34⁺ cells following transduction of FLT3 variants as in H. P-values were calculated using Student’s t-test compared to CTRL (mean + SD of n = 6 transductions). * P < 0.05, ** P < 0.01, **** P < 0.0001. See also Figure S5 and Table S3.

Figure 6.. Dysregulated transcriptional programs in malignant progenitors

A. Scatterplot positions genes (dots) by their preferential expression in malignant HSC/Prog-like cells relative to normal counterparts (x-axis), and by their correlation to HSC/Prog prediction scores across malignant cells (y-axis). Genes in the top right are preferentially expressed in malignant HSC/Prog-like cells, relative to normal progenitors and other malignant cell types (red). B. Heatmap shows expression of surface markers (rows) in normal BM cells (left, columns) or malignant cells from diagnostic AML aspirates (right, columns). CD14 is shown for comparison. P-values between HSC-like cells and normal HSCs are calculated by FDR-adjusted Kruskal test. C. Heatmaps show expression of normal BM-derived signature genes for HSC/Prog, GMP or differentiated myeloid cells (n = 90; rows) in normal BM (left, columns) or malignant AML cells (right, columns). Cells are ordered by their classifier prediction scores (shown on top). Cells that express cell cycle genes are indicated. Primitive AML cells co-express HSC/Prog and GMP programs. D. Plot shows correlation of 30 normal BM-derived HSC/Prog signature genes (red dots) with GMP prediction score across normal or malignant cells. Right: Plot shows correlation of 30 normal BM-derived GMP signature genes (blue dots) with HSC/Prog prediction score across normal or malignant cells. HSC/Prog genes and GMP genes are aberrantly correlated in malignant cells. P-values were calculated by paired Wilcoxon test. E. HSC/Prog-like and GMP-like signatures were applied to TCGA RNA-seq profiles. Heatmap shows expression of 60 signature genes (rows) across 179 bulk AML profiles (columns). F. Kaplan-Meier curves show the survival of 179 AML patients stratified by signature scores in E. Patients with higher HSC/Prog-like scores have worse outcomes. P-value was calculated by log-rank test. See also Figure S6 and Table S3–4.

Figure 7.. AML-derived monocyte-like cells have immunomodulatory properties

A. KNN visualization shows all T- and NK cells identified in normal BM and AML samples. BackSPIN analysis distinguished three clusters of cells that express markers of naïve T-cells, CTLs or NK cells. B. Boxplots show the relative numbers of cells annotated as T-cells or CTLs by scRNA-seq (median ± quartiles for 4 normal BMs and 16 diagnostic AMLs). C. Pie charts show relative numbers of cells annotated as CTLs or naïve T-cells by scRNA-seq (mean for two normal BM donors and six diagnostic AMLs with >50 T / NK cells). D. Representative IHC stains for T-cells (CD3) and CTLs (CD8) in normal BM and AML. H & E stains are also shown. Scale bar 50 μm. E. Boxplots show relative numbers of T-cells or CTLs identified in IHC stains (median ± quartiles for 15 normal BMs and 15 diagnostic AMLs). F. Pie charts show relative numbers of CTLs (CD8⁺), T-regs (CD25⁺FOXP3⁺) and other T-cells, per IHC stains (mean for 15 normal donors and 15 AMLs). AMLs have fewer T-cells and CTLs, but greater numbers of T-regs, consistent with an immunosuppressive microenvironment. G. Scatterplot shows 2,385 malignant monocyte-like cells from AMLs (red dots) and 567 monocytes from normal BMs (black dots). Cells are placed according to their signature scores for Mono1 (right), Mono2 (left), Mono3 (up) and Mono4 (down) (Villani et al., 2017). H. Barplot shows activation of a CD4⁺ T-cell line after stimulation with CD3/CD28 beads in vitro. T-cell activation was read out by an NFAT reporter. The assay was performed in the absence (Control) or presence of OCI-AML3 or MUTZ-3 cells (mean ± SD of n ≥ 3 experiments). I. Barplot shows activation of primary CD4⁺ T-cells after stimulation with CD3/CD28 beads in vitro. T-cell activation was read out by flow cytometry for CD69. The assay was performed in the absence or presence of MUTZ-3 cells (mean ± SD of n = 6 replicates). J. Barplots show activation of a CD4⁺ T-cell line as in H. The assay was performed in the presence of 100,000 sorted CD34⁺ or CD14⁺ MUTZ-3 cells (n = 3 experiments). K. Barplots show activation of a CD4⁺ T-cell line as in H, J. The assay was performed in the presence of 100,000 sorted CD14⁻ or CD14⁺ primary cells from normal BMs (n = 6 donors) or AML aspirates (n ≥ 3 technical replicates each). Significance is only indicated when T-cell activation was reduced >1.5-fold compared to Control. L. Heatmap shows expression of curated immunomodulatory genes (rows) in monocytes from normal BM (left, columns) and monocyte-like cells from AMLs (right, columns). Only AMLs with >50 monocyte-like cells are shown. M. Kaplan-Meier curves show the survival of 179 AML patients stratified by expression of MRC1/CD206 or CD163. P-values in M were calculated by log-rank test. All other P-values were calculated by Student’s t test. * P < 0.05, ** P < 0.01, *** P < 0.001, **** P < 0.0001. See also Figure S7 and Table S4.