Quantitative molecular cartography of emergency myelopoiesis reveals conserved modules of hematopoietic activation

SUMMARY

Hematopoietic stem and progenitor cells (HSPC) respond to infections, inflammation, and regenerative challenges using emergency myelopoiesis (EM) pathways to amplify myeloid cell production. However, it remains unclear how various EM inducers regulate HSPCs using shared or distinct molecular mechanisms. Here, we generate a comprehensive and generalizable cell annotation method (HemaScribe) and a refined quantitative model of hematopoietic differentiation (HemaScape) using single cell RNA sequencing (scRNA-seq) of murine HSPCs, which we apply to a broad range of EM modalities. We uncover multiple strategies to enhance myelopoiesis acting at different levels of the HSPC hierarchy, which are associated with both unique and shared transcriptional response modules. In particular, we identify a myeloid progenitor-based EM activation module across diverse inflammatory challenges that is conserved in humans and informs outcome in adult and pediatric acute myeloid leukemia. Our work illuminates fundamental regulatory mechanisms in hematopoietic regeneration that have direct translational applications in disease contexts.

eTOC BLURB

Swann et al. conducted comparative analysis of single cell RNA sequencing data from emergency myelopoiesis models, finding that different perturbations act at distinct levels of the hematopoietic hierarchy. They identify a conserved myeloid progenitor-based transcriptional activation module across multiple disease conditions, which informs outcome in human acute myeloid leukemia.

Graphical Abstract

INTRODUCTION

Hematopoiesis is highly adaptable to demand during infections, inflammation, cancer, and chemotherapy¹. In response to these stressors, hematopoietic stem and progenitor cells (HSPC) in the bone marrow (BM) niche recruit cellular and molecular mechanisms, collectively termed emergency myelopoiesis (EM) pathways, which coordinate rapid production of new myeloid cells². These EM pathways are normally terminated upon resolution of the initiating insult, but are persistently or inappropriately activated in aging, cancer, and chronic inflammation, rendering them important potential targets for therapy^1,2.

A wide range of mechanisms are important for murine EM pathway engagement, including increased proliferation³, altered cell death⁴, metabolic remodeling⁵, and accelerated myeloid differentiation^6,7. In hematopoietic stem cells (HSC), precocious activation of the myeloid transcription factor (TF) PU.1 by interleukin (IL)1β or tumor necrosis factor (TNF)-α drives enhanced myeloid commitment at the top of the HSPC hierarchy^8,9. In downstream multipotent progenitors (MPP), the lymphoid-biased MPP4 subset is reprogrammed towards a myeloid fate by IL6^[10,11], while the FcγR⁺ subset of myeloid-biased MPP3 undergoes accelerated differentiation toward the granulocyte macrophage progenitor (GMP) stage upon inflammatory stimulation¹². Downstream of MPPs, regenerative and inflammatory challenges induce transient self-renewal and expansion of GMPs, leading to emergence of GMP clusters that act as critical sites of neutrophil amplification¹³. Collectively, these examples illustrate how HSPC differentiation landscapes can be altered to enhance myelopoiesis and highlight the diversity of mechanisms available to the hematopoietic system to respond to increased demand for mature myeloid cells.

Despite this general understanding of EM pathways, it remains unclear whether all EM inducers alter activity at the same levels of the hematopoietic hierarchy, or if different perturbations exert focused effects on specific populations. Murine lineage tracing studies have recently suggested that HSC are largely dispensable for EM engagement during sepsis or inflammation, inferring that both steady state and regenerative myelopoiesis are solely dependent on MPPs and their progeny^14,15. It is also currently unknown whether each perturbation activates a unique set of molecular mechanisms in HSPCs or recruits shared effector modules. The former might be more effective in resolving the initial insult but would come at a high evolutionary cost of encoding information for each stimulus, while the latter represents a more efficient means to encode perturbation responses that could provide attractive therapeutic targets for mitigating maladaptive EM responses. However, there have been few attempts to search for shared activation modules across biological conditions, or to evaluate their possible clinical significance. Here, we leveraged the power of our extensive phenotypic and molecular characterization of 9 different perturbations driving EM engagement to uncover how different EM inducers remodel the HSPC hierarchy.

RESULTS

Consistent annotation of HSPCs in scRNA-seq datasets

To understand how different perturbations remodel the murine hematopoietic system, we first addressed two technical challenges in scRNA-seq analysis of HSPC datasets: (1) how to annotate the same cell types consistently and with high resolution, and (2) how to reconstruct the topology of normal hematopoietic differentiation and quantify its disruption. To overcome the first challenge, we created a new HSPC annotation approach called “HemaScribe” (Figure 1A). We first curated bulk RNA sequencing profiles for purified BM cell types^16–18 (Table S1A) to score cells according to predicted identity (broad classifier) (Figure 1B). This approach was highly effective for mature cells in whole BM datasets (Figure S1A) but did not provide sufficient resolution for HSPCs, particularly for known MPP and lineage-committed progenitor subsets^12,19–21. To remedy this, we created a reference dataset by FACS-sorting 10 different HSPC populations (i.e., HSC, short-term [ST]-HSC, MPP2, FcγR positive [FR⁺] and negative [FR⁻] MPP3, MPP4, common lymphoid progenitor [CLP], megakaryocyte progenitor [MkP], erythroid progenitor [EryP], and GMP) before labelling each cell type with different oligo-conjugated antibodies and pooling them for sequencing (Figures S1B-S1D). We used this dataset to train a classifier based on anchor integration and label transfer (fine classifier) (Figure 1C), which achieved high accuracy in reproducing the ground truth cell type represented by oligo hashtags (Figures 1D and 1E). The lowest recall was observed for the most transitional cell types, particularly for MPP2 and FR⁺ MPP3 that we recently described as intermediate between FR⁻ MPP3 and GMP¹². Accordingly, calculation of silhouette scores in either transcriptomic or flow cytometric data showed that MPP2 and FR⁺ MPP3 had the least distinct profiles (Figure S2A), explaining the difficulty in identifying these cells. Our classifier produced robust results when applied to bootstrap samples of the same FACS-sorted Lin⁻/c-Kit⁺ (LK) and Lin⁻/Sca-1⁺/c-Kit⁺ (LSK) scRNA-seq datasets, with the lowest similarity indices seen again for the most transitional cell types (Figure S2B). Next, we increased the annotation resolution of the GMP population using bulk RNA-seq references for subgroups of granulocyte progenitors (GP), common monocyte progenitors (cMoP), and multilineage GMP (mGMP) that generate either GP or cMoP²² as directly tested in liquid culture²³ (Figure S2C-S2E). We then applied the complete classifier to cells labelled as “HSPC” by the bulk classifier to produce our HemaScribe annotation (Figure 1F; Table S1B,C), which yielded predicted cell types in LK/LSK scRNA-seq datasets at similar frequencies to flow cytometry (Figures 1G; Figure S2F). Collectively, this work creates a new identification approach tailored specifically to murine HSPC annotation, with all functions available to the scientific community as the R package “HemaScribe” (Resource Availability).

Figure 1. HemaScribe provides consistent annotation of cell types.

(A) Outline for cell type classification using bulk and scRNA-seq references. HSPC: hematopoietic stem and progenitor cell.

(B) Uniform manifold and approximation projection (UMAP) showing annotation of 10X scRNA-seq dataset from Lin⁻/c-Kit⁺ (LK) and Lin⁻/Sca-1⁺/c-Kit⁺ (LSK) BM cells.

(C) UMAP showing 10 different sorted HSPC populations annotated with HemaScribe fine classifier. CLP: common lymphoid progenitor; EryP: erythroid progenitor; FR^−/+ MPP3: FcγR^−/+ multipotent progenitor 3; GMP: granulocyte macrophage progenitor; HSC: hematopoietic stem cell; MkP: megakaryocyte progenitor; MPP: multipotent progenitor; ST-HSC: short-term HSC.

(D) Heatmap showing percentage of hashtag labelled cells annotated as indicated by HemaScribe fine classifier, scaled to each population.

(E) Precision and recall for fine classifier in indicated cell types.

(F) UMAP showing annotation of LK/LSK scRNA-seq data using HemaScribe fine classifier and GMP subclassification. cMoP: common monocyte progenitor; GP: granulocyte progenitor; mGMP: multilineage GMP.

(G) Bar charts showing frequency of indicated cell types among LSK (left) or LK (right) cells, derived from either flow cytometry or annotation of scRNA-seq datasets using HemaScribe fine classifier. Points represent individual mice (flow cytometry) or biological replicates (scRNA-seq) with mean ± S.D.; P. values, two-way ANOVA with Sidak’s post hoc test.

See also Figures S1 and S2.

HemaScribe outperforms other methods for HSPC annotation

To test how the HemaScribe approach compared to existing scRNA-seq annotation packages, we applied 3 other methods with built-in cell type dictionaries (Azimuth²⁴, CellTypist²⁵, and scType²⁶) to our HSPC reference dataset (Figure 2A). Compared to the HemaScribe fine classifier, the other annotation packages lacked resolution to identify HSPC subsets (Figure 2B), with HemaScribe producing significantly greater correspondence to ground truth labels (Figure 2C). This was achieved with modest computational resources, with HemaScribe annotating 20,000 cells in less than 50 seconds on a standard laptop computer (Figure S3A). Even when providing our reference dataset to train other classifiers, only scANVI²⁷ performed similarly to HemaScribe, with other packages having significantly lower precision and recall averaged across cell types (Figures S3B and S3C). To establish whether our approach recaptured known biological features, we compared HemaScribe annotations to reference datasets generated by bulk RNA-seq or single cell CITE-seq approaches^12,28,29, observing high concordance among all resources (Figure S3D-S3F). Next, to illustrate how HemaScribe provides new insights, we re-annotated a published dataset of Lin⁻ BM cells³⁰ (Figure 2D), finding that true HSC were present only in cluster “HSC-1”, whereas “HSC-2” and “HSC-3” contained chiefly MPP, and that “intermediate myeloid progenitors” (IMP) were primarily cMoP (Figure 2E). Similarly, annotation of a seminal LK/LSK dataset³¹ (Figure S4A) allowed us to characterize an “immature progenitor” cluster as being composed of MPP3 and MPP4, and to identify many ST-HSC in the “HSC” cluster (Figures S4B and S4C). Finally, when applying the HemaScribe broad classifier to unfractionated BM cells in the Tabula muris repository³² (Figure S4D), we found that one cluster previously labelled as “macrophages” actually expressed genes characteristic of plasmacytoid dendritic cells (Figure S4E), illustrating how our extensive curation of cell type references also improved the accuracy of mature cell annotation. Collectively, this work provides a consistent approach to annotate BM scRNA-seq datasets, achieving unprecedented resolution in distinguishing HSPC subgroups, and permitting coordinated molecular analysis and prospective isolation of the same cell types.

Figure 2. HemaScribe outperforms other annotation methods for HSPC datasets.

(A) Outline showing approach to annotate with HemaScribe or other annotation packages.

(B) UMAP projections showing results of the same reference dataset annotated with indicated packages and respective naming conventions. BM: bone marrow; EMP: early myeloid progenitor; Eryth: erythroid lineage cell; ISG: interferon-stimulated gene; LMPP: lymphoid myeloid multipotent progenitor; Mk: megakaryocyte; Prog Mk: megakaryocyte progenitor.

(C) Comparison of clustering correspondence between indicated annotation packages and oligo hash labels in reference dataset. Points represent scRNA-seq replicates with mean ± S.D.; P. value, one-way ANOVA with Tukey’s post hoc test.

(D) UMAP projections showing annotation of lineage negative (Lin⁻) BM cells according to source publication (left) or with HemaScribe (right).

(E) Sankey plots showing composition of clusters defined in source publication against HemaScribe. IMP: intermediate myeloid progenitor.

See also Figures S3 and S4.

A unified topology for hematopoietic differentiation

To overcome the second challenge of understanding how HSPCs differentiate at steady state, we developed the “HemaScape” tree model to visualize the topology of hematopoiesis (Figure 3A). Using our reference oligo-hashed scRNA-seq dataset, we employed a dimensionality reduction method (elastic embedding [EE])³³ to capture both local and global continuous structure, before locating regions of high cell density taken to represent stable cell states using the DensityPath algorithm³⁴. To delineate connections among density clusters, we calculated minimum spanning trees to represent differentiation trajectories (Figure 3B), and we used the same density clusters as nodes for partition-based graph abstraction (PAGA) analysis³⁵ to estimate connectivity among nodes (Figure 3C; Table S2A). We combined the branching structure and connectivity values into a HemaScape differentiation tree that depicted the arrangement of major HSPC branches (Figure 3D), cross-referenced with HemaScribe annotations (Figure S5A). To understand how HemaScape performed against other methods for trajectory inference, we used the same reference data as input for Monocle3^[36] and Slingshot³⁷ to create alternative tree structures (Figure S5B), which we compared to a hierarchy inferred from rates of propagation of a fluorescent label originating in HSC across a time course in Hoxb5-Cre:R26LSL-tdT mice³⁸ (Figure 3E). To establish which topology most effectively captured the tree structure proposed by label propagation data, we explored the possible fates of cells in the query tree structures and then tested whether these fates were permitted in the label propagation tree. Doing so, we found that HemaScape produced significantly lower rates of forbidden cell fates compared to Monocle3 or Slingshot (Figure 3F). Next, we tested which model provided the best estimate of trajectory progression by calculating the inverse correlation between pseudotime values against a score of transcriptional variability taken as a measure of lineage commitment in the CytoTRACE2 package³⁹. Across multiple replicates, pseudotime values from either HemaScape or Slingshot showed significantly greater inverse correlation with CytoTRACE2 score than Monocle3 (Figure 3G). Moreover, HemaScape pseudotime showed significantly better correlation with two other independent measures of lineage differentiation: one estimated from fluorescent label propagation³⁸ and the other from tracking of clonal differentiation after transplantation of barcoded HSC⁴⁰ (Figure S5C). Finally, by comparing TF regulons enriched in different limbs emerging from key branchpoints using SCENIC⁴¹, we found that HemaScape effectively recaptured known determinants of commitment to erythropoiesis (GATA1)⁴², myelopoiesis (C/EBPα)⁴³, and lymphopoiesis (ETS1, TCF4)⁴⁴ (Figure S5D). Together, these data show that the HemaScape tree provides a refined model of hematopoietic differentiation that outperforms other approaches for trajectory inference.

Figure 3. HemaScape topological landscape for hematopoietic differentiation.

(A) Outline showing how DensityPath was used to identify density clusters connected by minimum spanning trees. Density clusters were used as input for partition-based graph abstraction (PAGA) analysis, and these measures were combined to infer the HemaScape tree.

(B) Density landscape showing density clusters (black points) and minimum spanning tree (black lines), annotated by cell type.

(C) Elastic embedding (EE) showing PAGA nodes derived from density clusters, with connectivity among nodes.

(D) HemaScape tree structure, with edge thickness determined by PAGA connectivity and nodes colored by HemaScribe composition.

(E) Tree structure proposed in previous publication of label propagation data, where edge thickness indicates differentiation flux and nodes are colored by HemaScribe composition.

(F) Results of random walk modeling for fates permitted for cells in tree structures defined using indicated packages, showing rates of forbidden fates in each model compared to tree shown in (E). HS: HemaScape; MN3: Monocle3; SS: Slingshot. Points represent independent scRNA-seq LK replicates with mean ± S.D.; P. value, one-way ANOVA with Tukey’s post hoc test.

(G) UMAP projections showing pseudotime values inferred from HemaScape (left) and CytoTRACE2 score of transcriptional diversity (middle) in steady state scRNA-seq LK data, with bar chart (right) showing inverse correlation between pseudotime derived from indicated packages against CytoTRACE2 score. Points represent independent scRNA-seq LK replicates with mean ± S.D.; P. value, one-way ANOVA with Tukey’s post hoc test.

(H) Simplified HemaScape tree structure (left) used for estimation of node size, proliferation, cell death, connectivity values, and differentiation rates. Bar charts (right) show calculated cell death rates, inferred proliferation rates, inferred differentiation rates, and Palantir entropy for indicated nodes. MegE: megakaryocyte/erythroid; Ly.: lymphoid. Points represent biological replicates of LK scRNA-seq datasets with mean ± S.D.

See also Figures S5 and S6.

To understand the cellular behavior of different nodes in HemaScape, we used a simplified version of our proposed tree structure that condensed smaller nodes (Figure 3H). We estimated cell death rates by direct measurement of ApoTracker green staining (Figures S5E) and proliferation rates by calibrating a gene expression score of cell cycle activity against HSPCs identified in the Fucci2 cell cycle reporter mouse line⁴⁵ (Figures S5F; Table S2B). Using these parameters combined with node connectivity values from PAGA and estimates for node sizes from their natural frequencies in steady state LK datasets, we implemented an iterative random walk approach to estimate the “differentiation rate” for each node. Doing so, we inferred higher proliferation and differentiation rates in downstream lineage-committed progenitor nodes compared to HSC/MPPs (Figure 3H; Table S2C), especially in the megakaryocyte/erythroid (MegE) branches, which aligns with patterns observed from label propagation data³⁸ (Figure 3E). Moreover, using Palantir⁴⁶ (Figures S6A; Table S2D), we confirmed a tendency for greater transcriptional entropy of immature HSC/MPPs compared to lineage-committed erythroid or myeloid progenitors (Figure 3H). This approach also allowed us to confirm the projected terminal state probabilities of different MPP subsets^21,47, with MPP2 showing the greatest bias towards MegE cells, FR⁺ MPP3 towards myeloid fate, and MPP4 towards lymphoid fate (Figure S6B). Finally, we sought to identify new surface markers enriching HemaScape nodes to facilitate future investigations. By mapping a published CITE-seq dataset⁴⁸ to HemaScape, we identified higher CD24 expression in nodes 1+2 and greater CD62L expression in node 9 (Figure S6C). We then isolated CD24⁺ LSK and CD62L⁺ LSK cells by flow cytometry, labelled them with oligo-conjugated antibodies, and pooled them for 10X scRNA-seq. Upon projection of the demultiplexed samples, we found that CD24 effectively enriched for nodes 1+2, as well as for EryP, and CD62L⁺ cells were found preferentially in node 9 (Figure S6D). Taken together, our findings outline a new quantitative model for hematopoietic differentiation that recapitulates numerous experimental observations, facilitates prospective isolation of cells at key differentiation nodes, and enables comparison to perturbed conditions. We provide functions to map query datasets to our EE embedding and assign HemaScape nodes in the R package “HemaScribe” (Resource Availability).

Different perturbations affect distinct levels of the hematopoietic system

We next compared scRNA-seq datasets from EM perturbation models, representing inflammation, regeneration, and responses to pathogen-derived products. Inflammation was induced by 7 (short-term) or 20 (chronic) daily injections of interleukin-1 (IL1)^8,49, regeneration by 2–4 daily injections of granulocyte colony stimulating factor (G-CSF) or occurring 8 days after administration of the cytotoxic drug 5-fluorouracil (5FU)¹³, and response to pathogen-derived products was studied 16 hours after lipopolysaccharide (LPS) injection²⁸ or 2–4 days after polyinosinic:polycytidylic acid (pIC) injection^50,51 (Table S2E). For comparison, we also included published datasets from mice exposed to erythropoietin (EPO) for 2 days⁵², and from embryonic day 18.5 (E18.5) fetal²⁸ and 24-month-old aged⁵³ mice. To evaluate quantitative changes in HSPC populations, we mapped LK scRNA-seq datasets to our reference embedding, applied HemaScribe annotations (Figure 4A; Table S1C), and compared cell frequencies using single cell differential composition analysis (scDC)⁵⁴. Importantly, we validated that HemaScribe performed well in perturbation conditions by comparing our steady state classifier with a similarly generated complementary algorithm trained on 5FU-treated HSPCs (Figure S6E). Consistent with previous findings¹³, 5FU caused massive expansion of HSC/MPP at the expense of lineage-committed progenitors (Figure S7A). In contrast, pIC and LPS caused more subtle changes in MPPs, with pIC also increasing the frequency of mGMP (Figure 4B; Figure S7B). Administration of G-CSF or IL1 had little impact on HSC/MPP frequency but caused dramatic expansion of GMP at the expense of EryP (Figure 4C; Figure S7C), whereas EPO had the opposite effect (Figure S7D). HemaScribe annotation of fetal and aged datasets also confirmed previous findings^28,53, with expanded EryP in E18.5 embryos, and a dramatically enlarged HSC compartment in old mice compared to 3-month-old adult controls (Figure S7E and S7F). Taken together, these results indicate that different EM perturbations and developmental stages produce distinct patterns of change in HSPC frequency.

Figure 4. Emergency myelopoiesis perturbations act at different levels of the hematopoietic hierarchy.

(A) Outline and key for HemaScribe annotation of perturbation datasets and matched controls. scDC: single cell differential composition.

(B-C) Stacked bar charts showing changes in frequency of indicated cell types in LK datasets upon (B) every 2-day injection of polyinosinic:polycytidylic acid (pIC) and (C) daily injection of granulocyte colony stimulating factor (G-CSF). P. values, scDC general linear models comparing day 4 (d4) to day 0 (d0).

(D) Simplified HemaScape tree (left) used as template for estimating node activities and heatmaps (right) showing transition matrices for optimal mass transport from indicated start nodes (y axis) to end nodes (x axis) from day 0 to day 4 of pIC and G-CSF time courses, estimated using Waddington-OT. Dashed boxes indicate author-selected areas of difference between perturbations.

(E) Outline and key for changes in PAGA connectivity compared to control datasets.

(F) MPP4 reprogramming with outline and PAGA subplot (left) showing changes in HemaScape node connectivity for indicated perturbations comparing day 4 to day 0, and line plot (right) showing change in connectivity between nodes O and H, with points representing scRNA-seq LK samples.

(G) Outline (left) showing transplantation of donor MPP4 into sublethally irradiated congenic recipients followed by injection of pIC or G-CSF to induce emergency myelopoiesis (EM). Arrows indicate injection times. Quantification (right) of proportion of donor-derived myeloid (Gr1⁺/CD11b⁺) cells in peripheral blood of recipients. Points represent n=5 mice per group with mean ± S.D.; P. values, multiple t-tests.

(H) Scatter plot showing log2 fold change (FC) in SCENIC transcription factor regulons comparing pIC at day 3 to control MPP4 (x axis) to differences between GMP and MPP4 (y axis).

See also Figures S6, S7, S8, S9, and S10.

To understand the cellular processes driving these quantitative changes, we recreated estimates of HemaScape node activity for each perturbation (Figure S7G; Table S2F-J), focusing on the effects of pIC and G-CSF. HemaScape predicted that pIC would broadly increase proliferation in MPP (C-F) and downstream myeloid progenitor nodes (G-J), whereas G-CSF would cause a focused increase in proliferation only in the mGMP-enriched node G (Figure S7H). These predictions were often mirrored by evaluation of the same nodes using Augur⁵⁵, which measures the strength of the transcriptional response to perturbation (Figure S7I). This revealed greatest transcriptional response in ST-HSC and MPP nodes (B-D) for pIC, whereas G-CSF had the greatest effect in myeloid progenitor nodes (H-I) (Figure S7J). To understand how these two perturbations affected differentiation of HSPCs across matching time courses, we used Waddington-OT⁵⁶ to generate an optimal transport map that defines a probabilistic mapping across time points based on transcriptional similarity, incorporating estimates of cell proliferation and death (Figure 4D). We inferred that early HSC/MPP made direct contributions to expansion of myelopoiesis upon pIC exposure, whereas G-CSF had little impact on this primitive compartment. This effect was apparent when evaluating the predicted fates of cells in the mixed MPP nodes C and E at day 0, both of which showed a bias towards myeloid fates at day 4 of pIC exposure (Figures S8A). To understand if the predicted differences between pIC and G-CSF were attributable to the capacity of different cells to respond to extrinsic signals, we evaluated expression of the cognate surface receptors toll-like receptor 3 (Tlr3), interferon (IFN) receptor α (Ifnar1), and G-CSF receptor (Csfr3) in scRNA-seq data and by flow cytometry (Figures S8B and S8C). From this, we also predicted aggregate receptor expression in each HemaScape node based on cell type composition (Figure S8D). These results showed that the principal receptor for type I IFNs was expressed broadly in the hematopoietic compartment, corresponding with widespread pIC-induced activation. While the highest G-CSF receptor expression was observed in myeloid-committed nodes, it was also expressed on most HSPCs, including early HSC/MPP that contributed little to myelopoiesis upon G-CSF injection. To understand whether differences in EM response were instead hardwired in epigenetic states, we exploited a combined LK/LSK multiome dataset²⁸, which we annotated using HemaScribe. We generated signatures based on the genes upregulated by each perturbation and scored each cell type for accessibility of those regions (Figure S9A). Strikingly, committed myeloid progenitor nodes (G-J) showed increased resting accessibility for genes upregulated by G-CSF compared to a random sample of genes, whereas MPP node E and MkP node K showed the greatest enrichment for the pIC signature (Figure S9B). Collectively, these data suggest that myeloid progenitors are epigenetically poised and competent to respond to G-CSF, which increases myeloid cell production without major contributions from upstream HSC/MPP, while MPP show the greatest poising to respond to pIC, which probably accounts for the broader range of EM transcriptional and differentiation responses in this context, including activation of early HSPCs.

Emergency myelopoiesis alters differentiation pathways

Next, we evaluated whether our HemaScape model could predict the engagement of new differentiation links² by calculating PAGA connectivity for each perturbation and its matched control sample and filtering for links showing changes of ≥0.15 (Figure 4E; Table S2F-J). Consistent with our optimal transport estimate results (Figure 4D), only pIC and LPS increased links between HSC and either MPP2 or mixed MPP nodes (Figure S9C). Moreover, we found that pIC and LPS increased connectivity between HSC and MkP nodes (Figure S9D), in line with the reported HSC-MkP bypass^57,58, for which pIC and downstream type I IFN responses are major drivers^59,60. To test this association, we isolated HSC from β-actin-Gfp mice⁶¹ and transplanted them into lethally irradiated recipients. Starting from the day after transplant, we injected recipients with either pIC or G-CSF according to the same dosage and schedule as for molecular analyses, finding that only pIC boosted the production of GFP⁺ platelets (Figure S9E). At the molecular level, SCENIC analysis of TF regulons in pIC-exposed HSC showed upregulation of factors that specify the megakaryocyte lineage, like Fli1 and Runx3^[62,63] (Figure S9F), suggesting pIC stabilizes the MkP bypass by causing precocious activation of platelet lineage-defining TFs. Finally, we evaluated the connectivity of MPP4-enriched nodes because this normally lymphoid-biased population can be reprogrammed by IL6^[10] or TF balance⁶ to produce myeloid cells upon demand. Among all perturbations, LPS, pIC, and IL1 enhanced connectivity from MPP4/CLP to mGMP/cMoP nodes, whereas G-CSF appeared to suppress this link (Figure 4F; Figure S9G). A similar effect was predicted for pIC using optimal transport methods, with cells in lymphoid node O showing considerable bias towards a myeloid fate at day 4 of pIC exposure, whereas G-CSF-treated cells remained committed to a lymphoid fate (Figure S9H). These changes in predicted fate were not associated with any differences in composition of the key MPP4 node D (Figure S9I), suggesting they were principally attributable to changes in cell behavior. To evaluate the same responses in vivo, we transplanted purified MPP4 into sub-lethally irradiated congenic recipients and started either pIC or G-CSF treatments the day after transplant. In line with our prediction, we found that only pIC prolonged the myeloid output of MPP4 (Figure 4G), and SCENIC analysis of TF regulons in pIC-exposed MPP4 again revealed upregulation of factors that can drive commitment of MPP to myeloid fate, like C/EBPα and C/EBPβ⁶ (Figure 4H). Accordingly, in silico knockouts of C/EBPα and C/EBPβ predicted much greater embedding shifts towards MPP in pIC-treated GMP than in G-CSF-treated cells (Figure S10A). Taken together, these results show that pathogen-derived products, especially pIC, can remodel the entire HSPC hierarchy through MPP reprograming and engagement of bypass mechanisms, whereas G-CSF acts only by enhancing myeloid progenitor-specific pathways that already operate at steady state (Figure S10B). Moreover, they illustrate the power of our quantitative HemaScape model in predicting the differentiation paths used by each EM inducer, creating a unique EM fingerprint delineated by activation of distinct molecular mechanisms at different levels of the HSPC hierarchy.

Shared and distinct molecular patterns of hematopoietic activation

Next, we asked whether common transcriptional modules could be activated by different EM stimuli to enhance myeloid output. To address this, we performed non-negative matrix factorization (NMF) on an integrated dataset containing 9 conditions and consisting of 25 scRNA-seq samples (Figure 5A; Table S3A). Cross-validation suggested a total of 40 factors was appropriate for our dataset (Figure S10C; Table S3B), which we categorized manually into those representing sources of technical or batch-related variation, cell type specific effects, or processes that varied by perturbation (Figure S10D; Table S3C). We also established a myeloid pseudotime gradient, beginning with HSC and progressing through density clusters to reach GP and cMoP (Figure S10E), which revealed a clear progression in molecular processes, starting with an HSC-associated murine factor (M7) driven by stemness factors like HOXA10, progressing through MPP-related factors (M9, M13) associated with HOXA7 and SOX4 TF regulons, and culminating with myeloid-related factors (M11, M23, M15, M30) associated with well-known myeloid TFs like C/EBPα, IRF8, C/EBPε, and GFI1 (Figure 5B). Among the factors that varied by perturbation, we selected four (M12, M18, M24, and M40) for further consideration owing to their interesting biological features. M18 was most active among HSC/ST-HSC, was specifically associated with acute IL1 exposure and both pathogen stimuli, LPS and pIC, and displayed the highest level of conservation across vertebrate species (Figure 5C; Figure S10F). Gene set enrichment analyses of its loading genes showed enrichment for pathways involved in tissue differentiation and TF activity (Figure S11), and SCENIC regulon scores indicated the greatest association with TCF12 and RUNX1 (Figure 5D). This suggests that M18 is associated with acute activation of early HSPCs, likely driven by RUNX1 and associated TFs, consistent with the role of RUNX1 in driving differentiation into multiple lineages⁶⁴ including myelopoiesis⁶⁵. While M12 was also most enriched in HSC/ST-HSC, it was instead activated by chronic IL1 and 5FU, was associated with expression of both JUN and FOS family members in its top loading genes, had strong correlations with the SCENIC regulons for AP-1 TFs, and was enriched for gene sets related to protein translation (Figure 5E and 5D; Figure S11). This suggests M12 is activated in chronic inflammatory settings and may promote myeloid differentiation in HSPCs, consistent with the role of JUN as an essential co-factor for PU.1 in myelopoiesis⁶⁶. M40 was also most enriched in HSC/ST-HSC and was strongly associated with LPS and pIC (Figure 5F). Accordingly, M40 contained many IFN-induced genes among its top loading genes, including Ly6a (Sca-1) and Irf7, which were enriched for gene sets related to viral infection, correlated with key IFN-related TFs like STAT1, STAT2, and IRF7 in SCENIC regulons, and enriched for genes induced by IFNs in immune response enrichment analysis (IREA)⁶⁷ (Figures 5D and 5F; Figure S11). Therefore, M40 explains the type I IFN response induced by LPS and pIC, which probably accounts for its more widespread activity in the hematopoietic system than either M12 or M18 (Figure 5G). In contrast, M24 was most active in myeloid progenitors like GMP, with activity associated with multiple perturbations including 5FU, G-CSF, pIC, and LPS (Figure 5H). Its top loading genes included metabolic genes (Ldha, Cox7b) and interferon-induced transmembrane proteins (Ifitm1, Ifitm3) normally more highly expressed in HSC than GMP²⁰, which were enriched for gene sets related to metabolic activation (Figure 5H; Figure S11). This suggests M24 is broadly associated with myeloid activation across perturbations.

Figure 5. Shared and distinct molecular signatures across emergency myelopoiesis perturbations.

(A) Outline showing how different EM datasets were integrated by canonical correlation analysis (CCA) before non-negative matrix factorization (NMF), finding k=40 factors.

(B) Succession of NMF factor activities across myeloid pseudotime, with top SCENIC regulons correlated with each factor. Most abundant cell types in pseudotime bins are shown in colored bar (HemaScribe annotation).

(C) Violin plots (left) showing activity of NMF factor M18 in combined HSC/ST-HSC (top) or GMP (bottom) across perturbations. Bar charts (right) showing loading scores for top 10 genes for each NMF factor. Colored violins indicate samples with greatest increases in factor score. Dots represent fold change of median values for perturbation groups over median values for control (Ctrl) values (dashed lines), with •: 1.5–2.0x, • •: 2.1–3.0x, and • • •:>3.0x.

(D) Bar charts showing top 5 Pearson correlation coefficients (coeff.) for NMF factor scores and SCENIC regulons for indicated factors, and results of immune response enrichment analysis (IREA) for M40. P. values, two-sided Wilcoxon rank sum test.

(E-F) Violin and bar charts for NMF factors (E) M12 and (F) M40 as outlined in (C).

(G) Change in relative importance of indicated factors across myeloid pseudotime.

(H) Violin and bar charts for NMF factor M24 as outlined in (C).

(I) Summary showing major cell types affected by different NMF factors. Myeloid activ.: myeloid activation.

See also Figures S10, S11, S12, and S13.

To investigate the stability of our NMF analysis, we conducted consensus (c)NMF as an alternative methodology⁶⁸, again using 40 factors named C1–40 (Figure S12A). By comparing the loading of genes in the w matrices, we observed a high level of correlation between models, with an average Spearman rho of 0.69 for best matching factors (Figure S12B). We noted that M24 showed the highest correlation with factor C13 in the cNMF model, which was itself also strongly correlated with factor M23. Indeed, our NMF model appeared to split C13 into two separate processes with greater biological resolution: myeloid differentiation (M23) and myeloid activation (M24) (Figure S12C). To establish the repeatability of the biological processes represented by our NMF factors, we tested whether they could also be observed in independently published resources, including acute IFN-α treatment over a time course of 72 hours⁶⁹, 2 hours after pIC or G-CSF injection⁷⁰, and following 5 weeks of chronic IL1 injection⁷¹ (Figure S13A; Table S3D). We confirmed that M40 was strongly associated with both type I IFN exposure driven by IFN-α or pIC, representing the most significantly upregulated factor in both settings (Figures S13A-S13C). Conversely, M12 was the second most upregulated factor with chronic IL1 and, to a lesser extent, with an independently generated time course of 5FU-mediated regeneration (Figure S13D). We also confirmed strong induction of M18 by pIC but not short-term IFN-α exposure (Figures S13B and S13C). Finally, M24 was more strongly induced in myeloid-committed progenitors than HSC in an independent G-CSF treatment dataset, as we observed in our own data (Figure S13C). Collectively, these results show that several distinct molecular effector modules can be activated in HSPCs by different groups of perturbations (Figure 5I) in a manner that is highly repeatable across independent studies and in ways that are likely to have important functional consequences for downstream EM responses.

A conserved EM-induced myeloid activation factor

The NMF factor M24 was increased in myeloid-committed progenitors across multiple conditions, making it an attractive core EM module for further study. In the context of 5FU exposure, M24 was among the most increased molecular processes at day 8, representing the time of maximal HSPC activation, in both myeloid-biased MPP3 and GMP (Figures 6A and 6B). Moreover, analysis of published bulk RNA-seq of BM GMP (Table S3E) from studies of murine myocardial infarction⁷², inflammatory colitis induced with dextran sodium sulfate (DSS)⁷³, inflammatory spondyloarthritis⁷⁴, and sepsis caused by LPS exposure⁷⁵ revealed significant enrichment for the top 100 loading genes of M24 in all settings (Figure 6C; Figure S14A), highlighting its commonality across inflammatory disease models. M24 top loading genes were significantly enriched for those induced by myeloid cytokines like granulocyte macrophage colony stimulating factor (GM-CSF) in IREA analyses (Figure 6D), as well as for genes associated with high output HSC⁷⁶ (Figure S14B), consistent with M24 being a driver of myeloid regeneration. Exploring possible upstream regulators, we found that M24 correlated highly with the SCENIC regulons for ERF, YBX1, and NFIA (Figure 6E). To understand these results in the wider context of molecular regulation of myelopoiesis, we performed in silico knockout screens of 33 TFs with known roles in myelopoiesis (Table S3F) using CellOracle⁷⁷, comparing day 8 5FU vs. day 20 IL1, which lacked increased M24 activity (Figure 6F). Constitutive myeloid factors like C/EBPα were equally important in both conditions, whereas PU.1, MYC, and IRF8 were more important for chronic IL1 exposure, and C/EBPβ, XBP1, and YBX1 for 5FU treatment. Based on these results, we identified the Y-box TF and cold shock protein YBX1 as a putative regulator of M24. YBX1 is a multifunctional protein regulating cell proliferation and differentiation, as well as binding mRNA to influence its stability, splicing, and translation⁷⁸. To determine whether YBX1 regulates genes in the M24 factor, we re-analyzed published cross-linking immunoprecipitation (CLIP)-sequencing data from a human MDA-231 breast cancer cell line⁷⁹. Interestingly, YBX1 was bound in the mRNA of many genes in the M24 module (Figure S14C), with significant overrepresentation of the top loading genes in YBX1 peaks (Fisher’s exact test, odds ratio 5.9, p=1.35×10⁻¹⁵) (Figure S14D). This suggests YBX1 might drive M24 by stabilizing RNA transcripts from activation-associated genes. To explore this, we first assessed YBX1 protein levels by flow cytometry in myeloid progenitors, confirming significant increases in day 8 5FU-treated MPP3, with no change in day 7 IL1-treated MPP3 (Figure 6G). We then performed ex vivo liquid culture of GMP isolated from either day 8 PBS and 5FU-treated or day 7 IL1-treated mice and expanded them in the presence or absence of the YBX1 inhibitor (YBX1i) SU056^[80] or PU.1 inhibitor (PU.1i) DB2313^[81] (Figure 6H). Consistent with M24 activation, 5FU-treated GMP expanded significantly faster than control GMP, while YBX1i exposure restored this expansion to the level of control GMP (Figure 6I). Importantly, YBX1i had no significant effect on either control or IL1-treated GMP, nor did PU.1i suppress the expansion of 5FU-treated GMP, demonstrating the specificity of this effect (Figure 6J and 6K). Mechanistically, YBX1 inhibition caused selective killing of 5FU-exposed GMP without affecting their proliferation rates (Figure 6L). Collectively, these data emphasize the broad activity of M24 in settings of myeloid regeneration, identify YBX1 as a putative regulator of this molecular process, and uncover a selective and targetable dependency of EM-activated myeloid progenitors on YBX1 activity.

Figure 6. Effector module of myeloid activation is shared across regenerative contexts.

(A-B) Post-5FU time course with violin plots (left) showing NMF factor M24 cell scores in MPP3 and/or GMP and bar charts (right) showing P. values derived from Wilcoxon rank sum test with Bonferroni correction comparing the activity of all factors between 5FU day 8 and control (Ctrl) day 0: (A) scRNA-seq dataset; and (B) published SMART-seq dataset. Dots represent fold change of median values for perturbation groups over median values for Ctrl values (dashed lines), with •: 1.5–2.0x, • •: 2.1–3.0x, and • • •:>3.0x.

(C) Enrichment plots for factor M24 top 100 loading genes in indicated published GMP bulk RNA-seq datasets. (N)ES: (normalized) enrichment score; P. values, permutation testing.

(D) Bar chart showing results of immune response enrichment analysis (IREA) for factor M24 genes. P. values, two-sided Wilcoxon rank sum test.

(E) Bar chart showing top 5 Pearson correlation coefficients (coeff.) for factor M24 scores and SCENIC regulons.

(F) Scatter plot showing scaled importance of transcription factors (TF) in myeloid differentiation of 5FU or IL1 treated cells, in a screen of 33 TFs using CellOracle. Values represent the inner product of perturbation score and development flow for each TF, scaled between 0 and 1 for each perturbation.

(G) YBX1 expression in cells isolated from mice treated once with PBS or 5FU (8 days post-injection) or daily with IL1 (after 7 days) with representative flow cytometry plot (left) of YBX1 intracellular staining in MPP3 and bar charts showing fold change in geometric mean fluorescence intensity (GeoMFI) for YBX1 compared to PBS in either MPP3 (middle) or GMP (right). No 1° Ab: no primary antibody. Points represent individual mice with mean ± S.D.; P. values, one-way ANOVA with post hoc Tukey’s test.

(H-K) Ex vivo TF inactivation in cultured GMP isolated from PBS, 5FU, or IL1-treated mice: (H) experimental scheme with GMP cultured with or without (±) inhibitors of YBX1 (SU056) or PU.1 (DB2313); and graphs showing the number of cells in culture over time following (I) YBX1 or (J) PU.1 inhibition in d8 5FU GMP, and (K) YBX1 inhibition in d7 IL1 GMP. Points represent n=4 biological replicates per group with mean ± S.D.; P. values, two-way ANOVA with Sidak’s post hoc test.

(L) Cellular response to YBX1 inhibition: GMP isolated from PBS or 5FU-treated mice were cultured ± SU056 and analyzed for proliferation (left) by flow cytometry for EdU incorporation after 1 hour (hr) and apoptosis (right) by caspase 3/7 (CC3/7) luminescent assay after 24 hours. Points show biological replicates with mean ± S.D.; P. values, two-way ANOVA with Sidak’s post hoc test.

See also Figure S14.

Myeloid activation factor predicts outcome in AML

We next used two complementary approaches to assess whether similar effector modules could regulate human hematopoiesis. First, we created ~50-gene mouse signatures (mSig) of our 4 EM-associated mouse NMF factors (M12, M18, M24, M40) by identifying top loading genes that were significantly upregulated in relevant perturbations and derived the equivalent human ortholog signatures (hSig) to map similar molecular process in human HSPCs (Figure 7A; Table S3G). Second, we searched for common modules in human EM contexts by integrating published BM scRNA-seq datasets from healthy volunteers exposed to G-CSF for 5 days⁸² and BM cells treated ex vivo with either LPS or the TLR1/2 agonist Pam3CSK4 for 4 days⁸³ (Figure 7A; Figure S14E; Table S3H), before mapping to a human reference embedding⁸⁴ and performing NMF with 30 factors (Table S3I). Remarkably, we identified a human factor H2 that was activated in GMP, but not in HSC/MPP, by both LPS and Pam3CSK4 (Figure 7B), and shared many key genes with the murine-derived mSig24 (Figure 7C). Formal comparison of murine against human NMF factors (Table S3J) revealed the strongest correlation of M24 with H2 (Figure S14F), with YBX1 also identified as a major predicted regulator for H2 (Figure S14G). This suggests that the molecular process represented by H2/M24 is conserved across mammal species, validating our approach of identifying fundamental EM signatures in murine datasets.

Figure 7. Myeloid activation factor is conserved in humans and predicts outcome in acute myeloid leukemia.

(A) Outline for generation of murine (mSig) and human (hSig) EM signatures, and human NMF model of EM perturbations.

(B) Violin plots showing activity of human NMF factor H2 in HSC/MPP (top) and GMP (bottom) in indicated perturbation conditions. P3C4: Pam3CSK4. Dots represent fold change of median values for perturbation groups over median values for control (Ctrl) values (dashed lines), with •: 1.5–2.0x.

(C) Bar plots showing top 15 genes by NMF loading coefficient for human H2 and mSig24, with colored bars indicating shared genes.

(D) Bar charts showing normalized enrichment score (NES) for mSig in bulk RNA-seq of GMP from indicated acute myeloid leukemia (AML) models. P. values, permutation testing.

(E) Bar charts showing scores derived by multiplying loading scores for genes in EM signatures with log transcripts per million (TPM)+1 in bulk RNA-seq samples from TCGA-LAML and Beat AML cohorts. Points show values for individual patients with mean ± S.D.; P. values, one-way ANOVA with Tukey’s post hoc test.

(F) Kaplan-Meier curves for overall survival for individuals without TP53 mutations with top and bottom 20% of hSig24 scores in TCGA-LAML and Beat AML cohorts. P. values, log rank tests.

(G) Forest plots showing coefficients for Cox proportional hazard models for overall survival in TP53 wild type patients in TCGA-LAML and Beat AML cohorts, including scores for hSig as covariates alongside age, sex, and previous malignancy or treatment. P. values, Cox regression.

(H) Bar charts showing likelihood ratios for predictive capacity of prognostic models incorporating indicated molecular scores, compared to baseline European Leukemia Network (ELN)2022 risk scores. P. values, ANOVA test for Cox model fits based on the log partial likelihood.

See also Figures S15 and S16.

Since SARS-CoV-2 (COVID-19) has been associated with aberrant EM responses, we next tested our signatures among genes upregulated in peripheral blood HSPCs of COVID-19 patients⁸⁵, finding significant enrichment for hSig24 and hSig40, whereas hSig12 was downregulated (Figure S14H). Since appropriation of molecular processes that drive activation and proliferation is a common strategy of neoplasia, we next investigated whether our EM signatures were enriched in AML. First, we found that mSig24 was significantly enriched in an AML mouse model originating from transformation of GMP with the MLL-AF9 fusion protein⁸⁶, but not in the Tet2^f/f:Vav1-iCre^+/−:Flt3^ITD/+ model that originates from transformed HSC/MPP⁸⁷ (Figure 7D; Table S3E). Using a classification system based on predicted cell of origin⁸⁸, we found that human patients in the TCGA-LAML⁸⁹ and Beat AML^90,91 cohorts had higher hSig24 scores in the ‘primitive’, ‘intermediate’, and ‘GMP’ groups compared to the ‘mature’ subtype (Figure 7E; Table S4A-S4B), consistent with the relative enrichment of M24 in murine GMP and MPP3. Next, we asked whether hSig24 enrichment might predict survival outcomes in AML patients using univariable analysis of different molecular alterations alongside hSig24. We found that both TP53 mutation status and hSig24 were significantly associated with worse overall survival after false discovery rate correction in both TCGA-LAML and Beat AML cohorts (Figure S15A-S15C), and that patients with TP53 mutations in the Beat AML cohort had higher hSig24 scores (Figure S15D; Table S4A-S4B). Since TP53 mutations are associated with worse prognosis in AML^89,92, we performed two analyses to assess the independence of the contributions of hSig24 from TP53 mutation status. First, we constructed multivariable Cox regression models in both cohorts, finding that hSig24 remained a significant independent predictor of overall survival. Second, to examine the prognostic value of hSig24 independent of TP53 mutation status, we removed all mutant TP53 samples and then compared overall survival of wild type (WT) TP53 patients with the highest (top 20%) and lowest (bottom 20%) hSig24 scores in the TCGA-LAML and Beat AML cohorts. We found that patients with the highest hSig24 scores had significantly poorer survival in both cohorts (Figure 7F). Using Cox proportional hazards models that accounted for age, sex, previous malignancy, and treatment history as covariates, we discovered an association of hSig24 with poorer survival among WT TP53 patients in both TCGA-LAML and Beat AML cohorts, whereas other EM scores did not predict outcome (Figure 7G). We also found that both hSig24 and hSig40 predicted outcome in the pediatric TARGET-AML cohort (Figure S16A), despite the major differences in molecular and clinical features observed between adult and pediatric AML⁹³. In contrast, hSig24 was not useful in predicting outcome in lymphoid malignancies or in most solid cancers (Figure S16B-S16C), though it was associated with overall survival in patients with lung adenocarcinoma in the TCGA-LUAD cohort⁹⁴, in which others have also reported myeloid signatures that inform outcome^95,96. Finally, by comparing likelihood ratios, we found that inclusion of hSig24 significantly improved model fit over the established ELN2022 clinical score⁹⁷, which is a proxy for mutation and cytogenetic profiles and classifies TP53-mutated AML as a separate disease entity, suggesting that the additional contribution of hSig24 cannot be explained by cytogenetic risk alone. Conversely, other gene scores⁹⁸, such as the APS⁹⁹, LSC17^[100], and Stem11^[101] scores, did not provide improvement over the ELN2022 score (Figure 7H). Collectively, these data show that our discovery of a conserved effector module of myeloid activation has direct clinical relevance owing to its ability to predict outcome beyond known factors like TP53 mutation in both pediatric and adult AML.

DISCUSSION

Using novel methods generated specifically for comprehensive analysis of HSPC perturbations, we show that pathogen-associated EM stimuli like LPS and pIC cause widespread changes in the hematopoietic system, whereas the regenerative cytokine G-CSF exerts a focused effect on myeloid-restricted progenitors. These observations are probably explained by multiple factors, including expression of surface receptors, epigenetic poising, and recruitment of downstream molecular processes. Functionally, these differences likely reflect the evolutionary importance attached to severe bacterial and viral infections, for which there is increasing evidence that the nature and extent of HSPC EM responses can dictate clinical outcome¹⁰². In contrast, G-CSF is primarily considered to have a role in homeostatic myeloid cell production¹⁰³.

Through integrative analysis of multiple perturbation conditions, we identify common EM modules recruited in non-overlapping combinations in different EM responses. This suggests that the hematopoietic system has only a limited number of ways to respond to diverse stimuli, which represents an efficient means to encode activation responses in the genome. We identify M24 as a shared activation module among myeloid progenitors, which also predicts outcome in human adult and pediatric AML, suggesting it is hijacked in cancer to produce more severe disease manifestations. Interestingly, YBX1, which we identify as a candidate driver of the myeloid activation factor, is an important dependency in AML cell lines¹⁰⁴, and the small molecule YBX1 inhibitor SU056 shows promise for its treatment in initial evaluations¹⁰⁵. Here, we find that YBX1 could have important physiological functions in myeloid cell expansion, which is consistent with its known roles in stabilizing transcripts of TFs that drive myelopoiesis, particularly MYC^104,106. Together, this suggests an evolutionary trade-off in maintaining molecular mechanisms that promote regeneration, but which could be exploited in mutated cancer cells and maladaptive conditions to permit uncontrolled expansion. Collectively, our work illuminates fundamental regulatory mechanisms in hematopoietic regeneration and identifies a common EM effector module that informs outcome in human disease contexts.

Limitations of the study

While we performed detailed comparisons of hematopoietic responses to pIC and G-CSF using harmonized time courses, our findings cannot be compared directly to other perturbations applied with different timings. Similarly, our conclusions on the differential effects of pIC and G-CSF might differ with more prolonged exposures, and our transcriptomic analyses have little power to separate direct from indirect effects of different perturbations. In this context, greater integration with other regulatory layers of cell activation, particularly chromatin accessibility and epigenetic modifications, will be needed to disentangle the direct and indirect effects of different perturbations.

STAR METHODS

Resource Availability

Lead Contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Emmanuelle Passegué, ep2828@cumc.columbia.edu.

Materials Availability

This study did not generate new unique reagents.

Data and code availability

Single cell RNA sequencing datasets generated in this study have been deposited at GEO. Accession numbers are listed in the Key Resources Table, along with details of all published datasets re-analyzed in this study. Functions for HemaScribe annotation of cells in scRNA-seq datasets and mapping to the HemaScape differentiation landscape are available in the R package “HemaScribe” with supporting reference data, available for download at github.com/RabadanLab/HemaScribe. The code for HemaScribe is archived at Zenodo, doi:10.5281/zenodo.19699029. All other analyses were conducted with publicly available software packages, as outlined in the Key Resources Table. Any additional information required to re-analyze the data reported in this paper is available from the lead contact upon request.

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rat anti mouse Ter119-PE/Cy5 (TER-119)	Invitrogen	Cat# 15-5921-82
Rat anti-mouse Ter119-PE/Cy5 (TER-119)	BioLegend	Cat# 116210
Rat anti mouse c-Kit-APC/Cy7 (2B8)	BioLegend	Cat# 105826
Rat anti-mouse Sca-1-BV421 (D7)	BioLegend	Cat# 108128
Rat anti-mouse Flt3-biotin, (A2F10)	BioLegend	Cat# 135308
Rat anti-mouse Flt3-PE (A2F10)	Invitrogen	Cat# 12-1351-82
Armenian hamster anti-mouse CD48-AF700 (HM48-1)	BioLegend	Cat# 103426
Rat anti-mouse CD41-BV510 (MWReg30)	BioLegend	Cat# 133923
Rat anti-mouse CD150-BV650 (TC15-12F12.2)	BioLegend	Cat# 115932
Rat anti-mouse CD127-PE (A7R34)	Invitrogen	Cat# 12-1271-82
Rat anti-mouse CD3e-PE/Cy5 (145-2C11)	BioLegend	Cat# 100310
Rat anti-mouse CD4-PE/Cy5 (RM4-5)	BioLegend	Cat# 100514
Rat anti-mouse CD5-PE/Cy5 (53-7.3)	BioLegend	Cat# 100610
Rat anti-mouse CD8a-PE/Cy5 (53-6.7)	BioLegend	Cat# 100710
Rat anti-mouse B220-PE/Cy5 (RA3-6B2)	BioLegend	Cat# 103210
Rat anti-mouse CD11b-PE/Cy5 (M1/70)	BioLegend	Cat# 101210
Rat anti-mouse Gr-1-PE/Cy5 (RB6-8C5)	BioLegend	Cat# 108410
Rat anti-mouse CD34-FITC (RAM34)	Invitrogen	Cat# 11-0341-85
Rat anti-mouse CD34-biotin (MEC14.7)	BioLegend	Cat# 119304
Rat anti-mouse Ly6G-FITC (1A8)	BioLegend	Cat# 127606
Rat anti-mouse Ly6C-APC (HK1.4)	BioLegend	Cat# 128016
Rat anti-mouse Ly6C-APC/Cy7 (HK1.4)	BioLegend	Cat# 128026
Rat anti-mouse CD11b-PE/Cy7 (M1/70)	Invitrogen	Cat# 25-0112-82
Rat anti-mouse CD115-PE (AFS98)	BioLegend	Cat# 135506
Rat mouse anti-mouse Ifnar-1-APC (MAR1-5A3)	BioLegend	Cat# 127313
Rat anti-mouse CD24-BV510 (M1/69)	BioLegend	Cat# 101831
Armenian hamster anti-mouse CD27-PE/Dazzle (LG.3A10)	BioLegend	Cat#124227
Rat anti-mouse CD102-AF488 (3C4)	BioLegend	Cat# 105609
Rat anti-mouse CD62L-APC (MEL-14)	Invitrogen	Cat# 17-0621-82
Armenian hamster anti-mouse CD61-PE (2C9.G2)	BioLegend	Cat# 104308
Rat anti-mouse CD150-APC (TC15-12F12.2)	BioLegend	Cat# 115909
Streptavidin-APC	BioLegend	Cat# 405207
TotalSeq B0301 anti-mouse hashtag 1	BioLegend	Cat# 155831
TotalSeq B0302 anti-mouse hashtag 2	BioLegend	Cat# 155833
TotalSeq B0303 anti-mouse hashtag 3	BioLegend	Cat# 155835
TotalSeq B0304 anti-mouse hashtag 4	BioLegend	Cat# 155837
TotalSeq B0305 anti-mouse hashtag 5	BioLegend	Cat# 155839
TotalSeq B0306 anti-mouse hashtag 6	BioLegend	Cat# 155841
TotalSeq B0307 anti-mouse hashtag 7	BioLegend	Cat# 155843
TotalSeq B0308 anti-mouse hashtag 8	BioLegend	Cat# 155845
TotalSeq B0309 anti-mouse hashtag 9	BioLegend	Cat# 155847
TotalSeq B0310 anti-mouse hashtag 10	BioLegend	Cat# 155849
Rabbit anti-mouse YBX1 (EP2708Y)	Abcam	Cat# ab76149
Rat anti-mouse TLR3-APC (11F8)	BioLegend	Cat# 141905
Rabbit anti-mouse CD114 (G-CSF-R) (723806)	ThermoFisher Scientific	Cat# MA5-24339
Alexa Fluor 647 goat anti-rat IgG (H+L)	Invitrogen	Cat# A21247
Alexa Fluor 488 goat anti-rabbit IgG (H+L)	Invitrogen	Cat# A11008
Bacterial and virus strains
Biological samples
Chemicals, peptides, and recombinant proteins
Hank’s buffered saline solution (HBSS) without calcium or magnesium	Gibco	Cat# 14175103
Phosphate buffered saline (PBS)	Gibco	Cat# 20012-027
Iscove’s modified DMEM medium (IMDM)	Gibco	Cat# 31980030
2-mercaptoethanol	Sigma-Aldrich	Cat# M6250
Fetal bovine serum (HI FBS)	Gibco	Cat# 16140-071
Penicillin/streptomycin	Gibco	Cat# 15-140-122
GlutaMAX supplement	Fisher Scientific	Cat# 35-050-061
Non-essential amino acids	ThermoFisher Scientific	Cat# 11-140-050
Sodium pyruvate	ThermoFisher Scientific	Cat# 11360070
Polyinosinic:polycytidylic acid (pIC)	Cytiva	Cat# 27-4732-01
5-fluorouracil	Sigma Aldrich	Cat# F6627
UltraPure 0.5M EDTA, pH8.0	Invitrogen	Cat# 15575-038
ApoTracker green	BioLegend	Cat# 427402
Murine recombinant SCF	PeproTech	Cat# 250-03
Murine recombinant IL-11	PeproTech	Cat# 220-11
Murine recombinant Flt3L	PeproTech	Cat# 250-31L
Murine recombinant IL-3	PeproTech	Cat# 212-13
Murine recombinant GM-CSF	PeproTech	Cat# 315-03
Murine recombinant G-CSF	PeproTech	Cat# 250-05
Human recombinant EPO	PeproTech	Cat# 100-64
Murine recombinant TPO	PeproTech	Cat# 315-14
SU056	Selleckchem	Cat# E1331
DB2313	Glixx Labs	Cat# 2170606-74-1
Critical commercial assays
Chromium Next GEM Single Cell 3’ Kit v3.1	10X Genomics	Cat# 1000268
Chromium Next GEM Chip G Single Cell Kit	10X Genomics	Cat# 1000127
Chromium GEM-X 3’ Kit v4	10X Genomics	Cat# 1000686
Chromium GEM-X 3’ Chip Kit v4	10X Genomics	Cat# 1000690
Dual Index Kit TT Set A	10X Genomics	Cat# 1000215
D5000 DNA Reagents	Agilent	Cat# 5067-5589
D5000 DNA TapeScreen	Agilent	Cat# 5067-5588
Foxp3/Transcription Factor staining buffer set	eBioscience	Cat# 00-5523-00
Cytofix/Cytoperm buffer	BD	Cat# 51-2090KZ
Perm/Wash buffer	BD	Cat# 51-2091KZ
Caspase-Glo 3/7 Assay System	Promega	Cat# G8090
Click-iT EdU Alexa Fluor 488 Flow Cytometry Assay Kit	Invitrogen	Cat# C10425
Deposited data
Single cell RNA sequencing data for short-term and chronic IL-1, 3 day pIC, days 8-14 5FU, hashing for 5FU-treated samples	This study	GEO: GSE296404
Single cell RNA sequencing data for G-CSF treatment and additional 5FU samples	This study	GEO: GSE298048
Single cell RNA sequencing data for hashing of steady state HSPC populations	Swann et al., 2026 (ref. 51)	GEO: GSE276784
Single cell RNA sequencing data for fetal E18.5 and LPS LK and LSK samples	Collins et al., 2024 (ref. 28)	GEO: GSE240915
Single cell RNA sequencing data for aged 24 month and steady state control samples	Mitchell et al., 2023 (ref. 53)	GEO: GSE169162
Single cell RNA sequencing with CITE-seq for HSPC populations	Klein et al., 2022 (ref. 29)	GEO: GSE145491
Single cell RNA sequencing with CITE-seq for HSPC populations	Ferchen et al., 2025 (ref. 48)	GEO: GSE266609
Single cell RNA sequencing for bone marrow populations	Tabula muris (ref. 32)	https://github.com/czbiohub-sf/tabula-muris
Single cell RNA sequencing for HSPC populations	Dahlin et al., 2018 (ref. 31)	GEO: GSE107727
Single cell RNA sequencing for EPO-treated samples and controls	Tusi et al., 2018 (ref. 52)	GEO: GSE89754
Single cell RNA sequencing for lineage negative bone marrow	Izzo et al., 2020 (ref. 30)	GEO: GSE124822
Single cell RNA sequencing for label propagation data	Kucinski et al., 2024 (ref. 38)	GEO: GSE207412
Single cell RNA sequencing for IFN treatment of bone marrow	Bouman et al., 2023 (ref. 69)	GEO: GSE226824
Single cell RNA sequencing for G-CSF and pIC treatment of bone marrow	Fast et al., 2021 (ref. 70)	GEO: GSE165844
Single cell RNA sequencing for chronic IL-1 treatment of bone marrow	McClatchy et al., 2023 (ref. 71)	GEO: GSE210026
Single cell RNA sequencing for human bone marrow exposed to LPS or Pam3CSK4	Reyes et al., 2020 (ref. 83)	https://singlecell.broadinstitute.org/single_cell/study/SCP550/an-immune-cell-signature-of-bacterial-sepsis-bone-marrow-stimulation
Single cell RNA sequencing for human bone marrow exposed to G-CSF	You et al., 2022 (ref. 82)	GEO: GSE193138
Single cell ATAC sequencing for steady state LK and LSK samples	Collins et al., 2024 (ref. 28)	GEO: GSE240915
SMART-seq for GMP with 5FU treatment	Hérault et al., 2017 (ref. 13)	GEO: GSE90799
CLIP-seq for MDA-231 cells	Goodarzi et al., 2015 (ref. 79)	GEO: GSE63605
Bulk RNA sequencing for HSC, MPP subsets, GMP	Collins et al., 2024 (ref. 28)	GEO: GSE240915
Bulk RNA sequencing for HSC, MPP subsets, GMP	Kang et al., 2023 (ref. 12)	GEO: GSE181902
Bulk RNA sequencing for GMP after myocardial infarction	Zhang et al., 2023 (ref. 72)	GEO: GSE169267
Bulk RNA sequencing for GMP with DSS colitis	Robles-Vera et al., 2025 (ref. 73)	GEO: GSE281376
Bulk RNA sequencing for GMP with spondyloarthritis	Regan-Komito et al., 2020 (ref. 74)	GEO: GSE126218
Bulk RNA sequencing for GMP with LPS exposure	Ikeda et al., 2023 (ref. 75)	GEO: GSE224102
Microarray for GMP transformed with MLL-AF9 retrovirus	Krivtsov et al., 2006 (ref. 86)	GEO: GSE18483
Bulk RNA sequencing for GMP from Vav-Cre+:Tet2f/f:Flt3ITD/+ mice	Shih et al., 2015 (ref. 87)	GEO: GSE57244
Signatures for high and low output HSC	Rodriguez-Fraticelli et al., 2020 (ref. 76)	Supplementary Table 2
Immune response enrichment analysis	Cui et al., 2024 (ref. 67)	https://www.immune-dictionary.org/app/home
TCGA-LAML cohort of AML patient data, mutation profiling, and RNA sequencing	cBioPortal; Genomic Data Commons Portal; Cancer Genome Atlas Research Network, 2013 (ref. 89)	https://www.cbioportal.org/study/summary?id=aml_tcga_gdc https://gdc.cancer.gov/about-data/publications/laml_2012
Beat AML cohort of AML patient data, mutation profiling, and and RNA sequencing	cBioPortal; BeatAML2; Bottomly et al., 2022 (ref. 91)	https://www.cbioportal.org/study/summary?id=aml_ohsu_2022 https://biodev.github.io/BeatAML2/
Target AML cohort of pediatric AML patient data and RNA sequencing	Genomic Data Commons Portal	https://www.cancer.gov/ccg/research/genome-sequencing/target https://portal.gdc.cancer.gov/projects/TARGET-AML
Other cancer datasets – TCGA-DLBC, BRCA	cBioPortal; Cancer Genome Atlas Network, 2012; 2014; Ciriello et al., 2015 (ref. 94, 141, 142)	https://www.cbioportal.org/study/summary?id=dlbclnos_tcga_gdc https://www.cbioportal.org/study/summary?id=brca_tcga_gdc
Experimental models: Cell lines
Experimental models: Organisms/strains
Mouse: WT-CD45.2: C57BL/6J	JAX	Cat# 000664
Mouse: BALB/cJ	JAX	Cat# 000651
Mouse: Fucci2a	Bertero & Vallier, 2015 (ref. 36)	N/A
Mouse: β-actin-GFP	Wright et al., 2001 (ref. 49)	N/A
Oligonucleotides
Recombinant DNA
Software and algorithms
HemaScribe	This study	https://github.com/RabadanLab/HemaScribe; Zenodo, doi:10.5281/zenodo.19699029
HemaScape	This study	https://github.com/RabadanLab/HemaScribe; Zenodo, doi:10.5281/zenodo.19699029
Interative website with HSPC populations	This study	http://3.233.5.190:8050
R (4.1.3)	The R Foundation	http://www.r-project.org
Python (3.9.20)	The Python Software Foundation	https://www.python.org
FlowJo (9/10)	BD	https://www.flowjo.com
Prism (9)	GraphPad	https://www.graphpad.com/features
Seurat (5.0.3)	Hao et al., 2021; 2024 (ref. 24, 109)	https://github.com/satijalab/seurat
Signac (1.14.0)	Stuart et al., 2021 (ref. 129)	https://stuartlab.org/signac/
CellOracle (0.20.0)	Kamimoto et al., 2023 (ref. 77)	https://morris-lab.github.io/CellOracle.documentation/
Palantir (1.4.1)	Setty et al., 2019 (ref. 46)	https://github.com/dpeerlab/Palantir
clusterProfiler (4.10.1)	Yu et al., 2012; Wu et al., 2021 (ref. 131, 132)	https://guangchuangyu.github.io/software/clusterProfiler/
DensityPath	Chen et al., 2019 (ref. 34)	https://github.com/ucasdp/DensityPath
pySCENIC (0.12.1)	Aibar et al., 2017 (ref. 41)	https://github.com/aertslab/pySCENIC
RcppML (0.3.7)	DeBruine et al., 2024 (ref. 130)	https://github.com/zdebruine/RcppML
cNMF (1.7)	Kotliar et al., 2019 (ref. 68)	https://github.com/dylkot/cNMF
PAGA	Wolf et al., 2019 (ref. 35)	https://github.com/theislab/paga
Azimuth	Hao et al., 2021 (ref. 24)	https://azimuth.hubmapconsortium.org/
CellTypist (1.7.1)	Xu et al., 2023 (ref. 25, 118)	https://github.com/Teichlab/celltypist
scType	Ianevski et al., 2022 (ref. 26)	https://sctype.app/
scANVI (1.4.1)	Xu et al., 2021 (ref. 27)	https://docs.scvi-tools.org/en/1.4.1/user_guide/models/scanvi.html
scDblFinder	Germain et al., 2022 (ref. 110)	https://github.com/plger/scDblFinder
UCell	Andreatta et al., 2021 (ref. 111)	https://github.com/carmonalab/UCell
SingleR	Aran et al., 2019 (ref. 112)	https://bioconductor.org/packages/release/bioc/html/SingleR.html
scikit-learn	Pedregosa et al., 2011 (ref. 122)	https://scikit-learn.org/
survival	Therneau, 2024; Therneau and Grambsch, 2020 (ref. 136, 137)	https://github.com/therneau/survival
survminer	Kassambara et al., 2024 (ref. 138)	https://github.com/kassambara/survminer
Waddington-OT	Schiebinger et al., 2019 (ref. 56)	https://broadinstitute.github.io/wot/
ComplexHeatmap	Gu, 2016 (ref. 135)	https://github.com/jokergoo/ComplexHeatmap
Other

Experimental Model and Study Participant Details

Animals

All animal experiments were conducted at the Columbia University Irving Medical Center (CUIMC) in accordance with approved Institutional Animal Care and Use Committee (IACUC) and in compliance with all relevant ethical regulations. Wild type (WT) C57BL/6J (CD45.2⁺) and B6.SJL-Ptprc^aPepc^b/BoyJ (CD45.1⁺) mice were purchased from the Jackson Laboratory and bred in house, as were the β-actin-Gfp mice⁶¹. Tg(Gt(ROSA)26Sor-Fucci2 mice⁴⁵ were kindly provided by Dr. H.L. Grimes (Cincinnati Children’s Hospital). Aged WT C57BL/6 mice were obtained from the National Institute on Aging (NIA) when they were 18 months old and used for experiments when they were 24 months old. Unless otherwise stated, mice were 8 to 12 weeks of age when used for experiments or as transplantation recipients. No specific randomization or blinding protocol was used with respect to the identity of experimental animals. Initial studies showed no difference in emergency myelopoiesis responses between sexes, and both male and female animals were used in all experiments. Animal facilities were maintained at 22 ± 2 °C and 50 ± 10% relative humidity on a 12–12 hours light–dark cycle, and mice were given ad libitum access to Purina LabDiet Rodent Feed and acidified water. Mice were euthanized by CO₂ asphyxiation followed by cervical dislocation.

Method Details

In vivo assays

5-fluorouracil (5FU) (Sigma Aldrich F6627) was dissolved in sterile PBS (10 mg/ml solution) and administered once at 150 mg/kg by intraperitoneal (i.p.) injection; control mice were injected with PBS. Recombinant human granulocyte colony stimulating factor (G-CSF) (Neupogen, Amgen) was diluted in sterile PBS (50 μg/ml solution) and injected i.p. daily (q24h) at 5 μg per mouse; control mice were injected with PBS. Recombinant murine interleukin (IL)1β (PeproTech) was reconstituted in PBS containing 0.1% bovine serum albumin (BSA) (5 μg/ml solution) and injected i.p. daily at 0.5 μg per mouse; control mice were injected with PBS/0.1% BSA. Polyinosinic:polycytidylic acid (pIC) (Cytiva) was dissolved in PBS (1.25 mg/ml solution) and injected i.p. every 48 hours (q48h) at 10 mg/kg; control mice were injected with PBS. Aliquots of pIC were warmed to 65°C for 15 minutes before injection. Recombinant murine erythropoietin (EPO, PeproTech) was reconstituted in PBS/0.1%BSA (10 U/ml solution) and injected i.p daily at 4 U/g i.p. for 2 consecutive days, as described⁵². Blood was collected by cardiac puncture after euthanasia and immediately transferred into EDTA-coated tubes (Greiner Bio-One, 0.5 ml) for complete blood count (CBC) analyses using a Genesis (Oxford Science) hematology analyzer.

Transplantation

For transplantation experiments, CD45.1 recipient mice were exposed to either 10.5 Gy lethal or 8 Gy sublethal irradiation delivered in split doses 3–4 h apart using an X-ray irradiator (MultiRad225, Precision X-Ray Irradiation) and injected retro-orbitally under isoflurane anesthesia with donor cells (500 GFP⁺ HSC or 5,000 CD45.2 MPP4) alongside 600,000 Sca-1-depleted CD45.1 helper BM cells. Starting the day after transplant, some recipients were injected with pIC (2 injections, 48 hours apart) or G-CSF (4 injections, every 24 hours), eventually followed by a second cycle of pIC (2 injections, 48 hours apart) or G-CSF (4 injections, every 24 hours) starting on day 21 after transplant. All recipient mice received water containing polymyxin (1000 U/ml, Sigma Aldrich P4932) and neomycin (1.2 mg/ml, Sigma Aldrich N1876) for 4 weeks following transplantation. For transplantation analyses, peripheral blood was collected with capillary tubes by retro-orbital bleeding under isoflurane anesthesia starting 7 days after transplant, and then every 4–5 days thereafter. For chimerism analyses, peripheral blood was collected into 1 mM EDTA containing ACK (150 mM NH₄Cl and 10 mM KHCO₃) red blood cell (RBC) lysis buffer for flow cytometry staining. For preparation of platelet-rich plasma, peripheral blood was collected into 1.5 ml tubes containing 20 μl 0.5M EDTA solution/pH 8 (Invitrogen 15575–038), centrifuged at 250 g for 10 mins, and transferred into a new 1.5 ml tube for flow cytometry staining.

Flow cytometry of hematopoietic cells

BM cells were obtained by crushing 8 bones (2 femurs, 2 tibiae, 2 humeri, and 2 pelvic bones) in staining medium composed of Hanks’ buffered saline solution without calcium or magnesium (HBSS, Gibco, 14175079) containing 2% heat-inactivated fetal bovine serum (FBS, Gibco HI FBS, 16140–071). RBCs were removed by lysis with ACK buffer (150 mM NH₄Cl and 10 mM KHCO₃), and single-cell suspensions of BM cells were purified on a Ficoll gradient (Histopaque 1119, Sigma-Aldrich). For HSC and progenitor isolation, BM cells were pre-enriched for c-Kit⁺ cells using c-Kit microbeads (Miltenyi Biotec, 130–091-224) and an AutoMACS cell separator (Miltenyi Biotec). Unfractionated or c-Kit-enriched BM cells were then incubated with the following antibodies: CD3e-PE/Cy5 (Invitrogen, 15–0031-63; 1:100), CD4-PE/Cy5 (Invitrogen, 15–0041-81; 1:1600), CD5-PE/Cy5 (BioLegend, 100610; 1:800), CD8a-PE/Cy5 (Invitrogen, 15–0081-81; 1:800), CD11b-PE/Cy5 (Invitrogen, 15–0112-81;1:1600), CD19-PE/Cy5 (BioLegend, 115510; 1:400), B220-PE/Cy5 (Invitrogen, 15–0452-81; 1:800), Gr1-PE/Cy5 (Invitrogen, 15–5931-82, 1:800), Ter119-PE/Cy5 (Invitrogen, 15–5921-83; 1:400), c-Kit-APC/Cy7 (BioLegend, 105826; 1:800); Sca-1-BV421 (BioLegend, 108128; 1:400); CD150-BV650 (BioLegend, 115931; 1:200); CD48-AF700 (BioLegend, 103426; 1:400); Flt3-PE (Invitrogen, 12–1351-82; 1:100) or Flt3-biotin (Invitrogen, 13–1351-85; 1:100) followed by streptavidin (SA)-APC (BioLegend, 405207; 1:200), CD127-PE (Invitrogen, 12–1271-82), CD34-FITC (Invitrogen, 11–0341-85; 1:50) or CD34-biotin (BioLegend, 119304; 1:100), CD16/32-PE/Cy7 (BioLegend, 101317; 1:800), CD41-BV510 (BioLegend, 133923; 1:200), and CD105-BV786 (BD, 564746; 1:200). For analysis of Fucci2 cell cycle reporter mice, CD34-biotin was used followed by SA-APC, Geminin-mVenus was detected in the B530 channel, and Cdt1-mCherry was detected in the Y615 channel. For analysis of surface receptor expression, IFNAR-1-APC (BioLegend, 127313, 1:200) was added to the panel. For G-CSF receptor (CD114), cells were first incubated with an unconjugated anti-CD114 antibody (0.25 μg per 10⁶ cells, ThermoFisher Scientific, MA5–24339) for 30 minutes, then washed and stained with goat anti-rat IgG(H+L) AF647-conjugated secondary antibody for 30 minutes (Invitrogen, A21247, 1:400). After washing, cells were then stained with other surface antibodies. For analysis of possible node markers, antibodies for CD27-PE/Dazzle (BioLegend, 124227, 1:200), CD102-AF488 (BioLegend, 105609, 1:200), CD62L-APC (eBioscience, 17–0621-81, 1:200), and CD24-BV510 (BioLegend, 101831, 1:400) were added to the panel. For analysis of apoptosis, cells were washed after initial staining then incubated with ApoTracker Green (BioLegend, 427401) according to the manufacturer’s instructions. For peripheral blood chimerism analyses after MPP4 transplant, cells were stained with CD11b-PE/Cy7 (Invitrogen, 25–0112-82; 1:800), Gr1-e450 (Invitrogen, 57–5931-82; 1:400), B220-APC/Cy7 (Invitrogen, 47–0452-82; 1:800), CD3-APC (Invitrogen, 17–0032-82; 1:100) and Ter119-PE/Cy5 (Invitrogen, 15–5921-83; 1:400) together with CD45.2-FITC (Invitrogen, 11–0454-85, 1:400) and CD45.1-PE (Invitrogen, 12–0453-83, 1:400). For peripheral blood chimerism analysis after GFP⁺ HSC transplant, platelet rich plasma was stained with Ter119-PE/Cy5 (Invitrogen, 15–5921-83; 1:400), CD61-PE (BioLegend, 104308, 1:400), and CD150-APC (BioLegend, 115910, 1:400), with acquisition also in the GFP channel. For short-term culture of GMP, GP and cMoP, cultured cells were stained with F4/80-APC (Invitrogen, 17–4801-82, 1:500), CD11b-PE/Cy7 (Invitrogen, 25–0112-82, 1:800), CD115-PE (BioLegend, 135506, 1:100), Ly6C-APC/Cy7 (BioLegend, 128026, 1:400), and Ly6G-FITC (BioLegend, 127606, 1:400). Stained cells were resuspended in staining medium containing 1 μg/ml propidium iodide (Sigma, P4170) for dead cell exclusion. Isolation of specific cell types was performed on a FACS Aria II SORP (BD) using double sorting for purity. Flow cytometric analyses were performed on a NovoCyte Quanteon or NovoCyte Penteon cell analyzer (Agilent). Data collection was performed using FACSDiva (v9) or NovoExpress (v1.6.2), and analysis was conducted using FlowJo (v9/v10).

Intracellular flow cytometry

For intracellular staining of YBX1, c-Kit-enriched BM cells were stained for surface markers for 30 minutes at 4°C (CD150-APC [BioLegend 115909, 1:400], CD48-AF700 [BioLegend, 103426; 1:400], c-Kit-APC/Cy7 [BioLegend, 105826; 1:800], Sca-1-Pacific blue [BioLegend 108120, 1:200], Flt3-PE [Invitrogen 12–1351-82, 1:100], CD16/32-PE/Cy7 [BioLegend, 101317; 1:800], and lineage stains in PE/Cy5 as described above) then fixed and permeabilized using the FOXP3/Transcription Factor Fixation/Permeabilization and 10X Permeabilization solutions (ThermoFisher Scientific, 00–5523-00) according to the manufacturer’s instructions. After this, cells were stained with rabbit anti-mouse YBX1 antibody (Abcam, ab76149, 1:200 dilution in 1X Permeabilization solution) for 1 hour at 4°C. After washing, cells were stained with goat anti-rabbit IgG AF488 (Invitrogen, A11008, 1:400 dilution) for 30 minutes at 4°C. Cells were then washed and re-suspended in 1X Permeabilization buffer for acquisition on a NovoCyte Penteon. EdU incorporation was measured using the Click-It EdU Flow Cytometry Assay (Invitrogen, C10425) kit according to the manufacturer’s instructions and AF488 azide, with readout on a NovoCyte Penteon.

Liquid culture of myeloid progenitors

For GMP cultures, cells were first sorted into 1.5 ml tubes containing HBSS with 2% FBS using a FACS Aria II (BD) on “yield” setting and then re-sorted on “4-way purity” to deposit 250 cells per well of a 96 well U bottom plate filled with 200 μl of Iscove’s modified Dulbecco medium (IMDM, Invitrogen 31–980-097) containing 5% FBS, 100 U/ml penicillin and 100 μg/ml streptomycin (Thermo Scientific, 15140122), 0.1 mM non-essential amino acids (Fisher Scientific, 11–140-050), 1 mM sodium pyruvate (Fisher Scientific, 11–360-070), 2 mM L-glutamine (Fisher Scientific 35–050-061), 50 μM 2-mercaptoethanol (Sigma, M7522), and the following cytokines (all from PeproTech): IL3 (10 ng/ml), granulocyte–macrophage colony-stimulating factor (GM-CSF; 10 ng/ml), stem cell factor (SCF; 25 ng/ml), IL-11 (25 ng/ml), Flt3L (25 ng/ml), thrombopoietin (TPO; 25 ng/ml), and erythropoietin (EPO; 4 U/ml). In some wells, the YBX1 inhibitor SU056 (Selleck Chemicals, E1331) was added at 500 nM, or the PU.1 inhibitor (Glixx Labs, 2170606–74-1) at 50 nM. Cells were counted every 2 days by removing 100 μl of medium and mixing with 50 μl of HBSS containing 2% FBS and 1 μg/ml propidium iodide before counting on a NovoCyte Penteon using the “absolute count” setting. For caspase 3/7 assays, 1,000 GMP were re-sorted on 4-way purity into wells of a 96 well plate and cultured as described above. After 24 hours, cells were resuspended by pipetting up and down before 40 μl was transferred to a white 384 well plate in triplicate. Then, 40 μl of Caspase-Glo reagent (Promega, G8090) was added to each well, and the plate was incubated for 1 hour in the dark at room temperature. After this, luminescence was detected using a SpectraMax iD3 plate reader (Molecular Devices) with SoftMax Pro 7.1 software. For measurement of EdU incorporation, 2,000 GMP were re-sorted on 4-way purity into wells of a 96 well plate and cultured as described above. After 1 hour, EdU was added to cultures at a final concentration of 10 μM and incubated for a further hour before flow cytometry analysis. For short-term cultures of GMP and GP/cMoP subpopulations, cells were cultured for 2 days as described above, except that GM-CSF was replaced with G-CSF (10 ng/ml, PeproTech) and analyzed by flow cytometry staining.

Generation of scRNA-seq datasets

Between 30,000 to 50,000 LK or LSK cells were sorted into 1.5 ml tubes containing a 1:1 mixture of HBSS and FBS and then rested for 1 hour on ice before pelleting at 350 × g for 5 minutes at 4°C. For sequencing of defined HSPCs, 8,000 cells of 10 different populations (HSC, ST-HSC, MPP2, FcγR⁻ MPP3, FcγR⁺ MPP3, MPP4, GMP, MkP, CLP, EryP) were each sorted in separate 1.5 ml tubes and labelled with 0.5 μg of oligo-conjugated antibodies specific for MHC class I and CD45 (BioLegend, TotalSeq B1–10) for 30 minutes on ice before washing three times with HBSS containing 2% FBS, and then pooled together for subsequent partitioning. For single or pooled samples, the supernatant was removed after centrifugation down to a volume of 40 μl before GEM generation and 3' RNA library preparation were performed according to 10X Genomics protocols CG000315 Rev E and CG000731, targeting 5000 cell data recovery. RNA libraries were pooled 1:1:1:etc., sequenced on an Illumina NovaSeq 6000 (2×150 bp, 2.5 billion reads), and aligned using Cell Ranger (v.7.0.1) to mouse genome mm10. Library concentrations and fragment sizes were evaluated using Qubit dsDNA HS assay kit (ThermoFisher Scientific) and TapeStation D5000 DNA ScreenTape analysis (Agilent).

Quality control for scRNA-seq data

Standard quality control filtering was performed on each newly generated scRNA-seq dataset, removing empty droplets (cells expressing fewer than 200 unique counts) and low-quality cells based on outliers for total counts detected (in log scale), number of unique genes detected (in log scale), percentage of counts in top 50% of expressed genes, and percentage of mitochondrial counts¹⁰⁷. Thresholds were automatically set at ±5 median absolute deviations (MADs) for each quality control variable. Additionally, cells with >8% mitochondrial counts were also excluded. The oligo-hashing reference samples (n=4) were demultiplexed using MULTI-seq¹⁰⁸ as implemented in Seurat v5^24,109, following which doublets and negative cells were removed and singlet cells were assigned to one of the hashing-labelled cell types. For all other non-hashing datasets generated in this study, doublet removal was performed using scDblFinder¹¹⁰.

HemaScribe: murine BM cell type annotations

We developed HemaScribe, a hierarchical strategy to annotate and classify HSPC subpopulations within the mouse BM by integrating multiple well-established methods for dissecting the identities of single cells with bulk and single cell transcriptomic references into one pipeline. HemaScribe consists of three modules:

The first module is a hematopoietic cell filter to exclude non-hematopoietic cells. It is based on gene set scores calculated using the UCell::ScoreSignatures_UCell function¹¹¹ with a list of hematopoietic marker genes that were selected from genes differentially expressed between all hematopoietic and non-hematopoietic cells in the Tabula muris droplet datasets, manually supplemented with additional genes for erythropoiesis (Ptprc, Itgb2, Cd79a, S100a8, Hba-a1, Pf4, Plac8, Ly6d, Ly6c2, Epor, Cd79b, Kel, Coro1a, Rhd, Gypa, Ctsg, Mpo). Depending on tissue source and contamination a sample may contain non-hematopoietic cells, and it is important to remove such cells to improve the efficiency of our subsequent classifiers which are trained solely on hematopoietic cell types. Therefore, after scoring, cells with a hematopoietic score below a manual threshold can be excluded from downstream analysis, and we found that a threshold of ~0.1 was appropriate for most BM datasets.

The second module is a broad classifier where cells are classified into broad BM populations using SingleR¹¹² based on their correlation with a curated set of bulk RNA-seq references^16–18 (Table S1). This was principally done to extract a subpopulation of BM cells corresponding to the early progenitors that align with the single cell LK and LSK reference datasets in the next step but is also useful for resolving other cell type populations (e.g., GMP and GP) in the marrow.

The third module is a fine classifier whereby cells that are identified as “HSPCs” by the broad-level classifier are assigned a finer cell type using Seurat’s integrative analysis based on anchor integration¹¹³ with our newly generated oligo-hashing reference samples. These reference datasets were created by separately concatenating the WT samples and processed using the Seurat functions NormalizeData, FindVariableFeatures, FindIntegrationAnchors, and IntegrateData with 3000 variable features and 3000 anchor features each. The integration step was performed with the Seurat commands FindTransferAnchors and TransferData using 30 principal components. After integrating our query dataset with the reference, the oligo-hashed labels are transferred onto each cell in the query. For ease of interpretation, we grouped together overlapping cell type labels in the broad-level and fine-level annotations to obtain a final combined annotation. We also developed a variant of HemaScribe using the 5FU reference hashing label datasets instead of the WT datasets following the same procedure.

To assess the performance of the fine classifier, we computed the per-cell type precision and recall on the training datasets as well as on held-out 5FU datasets. We also validated our HemaScribe annotations against previously published hematopoiesis datasets, including a recent transcriptomic atlas of mouse Lin⁻ BM cells³⁰ sequenced using the Chromium platform from 10X Genomics, a dataset of LSK cells subjected to CITE-seq²⁹, a collection of bulk RNA-seq samples from purified HSPCs^12,28, a scRNA-seq reference dataset for LK/LSK cells³¹, and a sample of unfractionated BM cells sequenced by 10X scRNA-seq³². For the CITE-seq data, a value of 1 was added to the antibody-derived tag (ADT) data, which were then normalized using the “CLR” option in the NormalizeData function in Seurat, with margin=2. Cutoff limits were applied for Flk2 (CD135), CD48, and CD150 by visual inspection, with the following values: MPP4: CD135>0.6 and CD150<1; MPP3: CD135<0.5 and CD150<1.5 and CD48>2.5; MPP2: CD135<0.5 and CD150>1.5 and CD48>2.5; HSC: CD135<0.5 and CD150>1.5 and CD48<2.5; and ST-HSC: CD135<0.5 and CD150<1.5 and CD48<2.5. To compare the results of HSPC-sorted cells with HemaScribe, we calculated the adjusted Rand index (ARI) for cell assignments using the adjustedRandIndex function from the R package mclust (version 6.1.2). To evaluate associations between HemaScribe and bulk RNA-seq samples, we first obtained pseudobulk profiles for each HemaScribe population from a control dataset incorporating 3 steady state WT LK and 3 LSK datasets using the AggregrateExpression function in Seurat, and we then calculated correlations between these profiles and bulk RNA-seq samples using the clustify function from the R package ClustifyR (version 1.22.0), using the top 2,000 highly variable genes from the scRNA-seq dataset. Next, we scaled each bulk RNA-seq sample and identified its greatest correlate among the HemaScribe populations, and we used these results to calculate the overall ARI for cell types represented in both datasets. In addition to testing the accuracy of our annotations on these external benchmarks, we also examined the internal consistency of our annotations by adapting the cluster stability measures from scclusteval¹¹⁴. Briefly, we repeatedly sub-sampled 80% of our control datasets 100 times, re-ran HemaScribe on each sub-sample, and recorded the cell type annotations from each run. Next, we computed the Jaccard similarity for each imputed subgroup in the subsampled dataset with the corresponding subpopulation imputed from the full dataset. A high median Jaccard similarity across all the runs for a given cell type means that the annotations for that cell type are robust and stable. We also computed the Jaccard similarity for each subgroup with the off-target subpopulations in the full dataset. A low median Jaccard similarity here means that discordant cell type assignments are uncommon. Finally, we ran HemaScribe on a variety of normal and perturbed hematopoiesis datasets (Table S1B). After filtering for low quality cells, all datasets generated in this study were preprocessed using the Seurat commands NormalizeData and FindVariableFeatures with default parameters before HemaScribe functions were applied. The previously published HSPC CITE-seq dataset²⁹, EPO dataset⁵², Tabula muris datasets³², and mouse Lin⁻ BM dataset³⁰ were re-preprocessed using the same methods, except that the list of included cells was retrieved from the original analyses. For all other external datasets, we used the original pre-processed data from the authors.

Comparison with other annotation methods

We benchmarked HemaScribe against five other commonly used annotation pipelines for scRNA-seq data. Each method was tested on oligo-hashing reference samples generated in this study, for which ground truth labels were extracted from the ADT tags as described above, and compared using precision and recall accuracy metrics. In general, we sought to test these methods using their default parameter settings and following the package vignettes and tutorials if provided. For Azimuth²⁴, we ran Azimuth::RunAzimuth function from the Azimuth R package (version 0.5.0) with reference=”bonemarrowref”, a human BM reference^115–117 provided as part of the package. For CellTypist^25,118, we used the celltypist.annotate function from the celltypist Python package (version 1.7.1) with the homology-converted combined adult human BM reference^118–121 from their organ atlas. For scType²⁶, we downloaded R scripts from https://github.com/IanevskiAleksandr/sc-type and ran run_sctype with known_tissue_type=“Immune system”. The correspondence of each of these annotations and for the HemaScribe fine classifier with ADT tags was evaluated by calculation of the adjusted Rand index using the function adjustedRandIndex in the R package mclust. Separately, we repeated this benchmarking analysis using custom references generated from our oligo-hashing reference samples, using leave-one-out cross-validation. Specifically, for each of the four oligo-hashing reference samples, we derived references from the other three datasets, trained models on them, and assessed their accuracy on the fourth left-out dataset in turn. This allowed us to use annotation methods that did not come with their own bundled reference data, such as scANVI²⁷ from scVI Python package (version 1.3.1). Following their tutorial, after the data was prepared we constructed a scVI model with scvi.model.SCVI (use_layer_norm=’both’, use_batch_norm=’none’, encode_covariates=True, dropout_rate=0.2, n_layers=2) and trained the model with scvi.model.SCVI.train for 400 epochs. Then, cell type labels were provided from the training samples and we initialized a scANVI model from the trained scVI model and further fine-tuned the model using scvi.model.SCANVI.train with training parameters max_epochs=20, n_samples_per_label=100. Finally, we predicted the cell labels of the held-out dataset by querying the model with scvi.model.SCANVI.train and max_epochs=100. We also assessed SingleR¹¹² in this way, using the command SingleR::SingleR with de.method=”wilcox” following a recommendation for using single-cell references in SingleR. From these results and for the HemaScribe fine classifier, we derived measures of precision and recall compared to ADT labels, separately for each cell type and also averaged across cell types. Averaged precision and recall values for each hashing dataset (n=4) were compared among annotation methods using one-way ANOVA with Dunnett’s post hoc test. Additionally, we profiled the runtimes of each method (HemaScribe, Azimuth, CellTypist, and scType) on an omnibus scRNA-seq dataset obtained by merging 9 LK/LSK datasets generated by us (EA001, AA003, EJ001, EO035, EO036, EO010, EO011, EJ007, EJ008 in Table S1B). From this dataset, we subsampled datasets of sizes 4000, 8000, 12000, 16000, 20000, 24000, 28000, 32000, 36000, and 40000, and measured how long each method took to run on the subsampled dataset. This was repeated ten times for each method and dataset size. For methods written in R, we took the “elapsed” time from system.time(); for methods written in Python, we used timeit.timeit(). For HemaScribe, we skipped the hematopoietic scoring module as it is not used for cell type annotation. These timing experiments were done on a 2019 MacBook Pro with a 2.6 GHz 6-Core Intel Core i7 processor and 16 GB memory running macOS 15.1.1.

Intrinsic separability of HSPC subsets in transcriptomic and flow-cytometric space

To assess the extent to which HSPC subpopulations are distinct from one another in transcriptomic landscape and thereby quantify how easily they can be computationally separated independent of any specific pipeline, we computed the silhouette scores based on the first ten principal components in scRNA-seq data against the ground truth hashing labels, using the function sklearn.metrics.silhouette_samples¹²². This gives a per-cell measurement of how close a cell is to the other cells within the same subpopulation, relative to cells from other subpopulations. Likewise, we also computed silhouette scores based on compensated, biexponentially transformed fluorescence data that were exported from FlowJo for n=3 replicates of flow cytometric profiles of the same cell populations, using the provided gating scheme (Fig. S1B).

Evaluation of cell frequency changes

To evaluate the statistical significance of changes in HSPC frequency annotated by HemaScribe in perturbation and matched control datasets, we used functions from the R package scDC, as outlined by the authors⁵⁴. First, we performed differential composition analysis using scDC_noClustering with 500 bootstraps, using the fine annotation from HemaScribe, in which we subdivided GMPs into constituent mGMP, GP, and cMoP. Next, we used fitGLM to fit generalized linear models with the CLP cell type as the reference category, and we took p-values from the fixed effect pooled results.

HemaScape: Trajectory analysis of hematopoiesis using DensityPath

To investigate the mechanisms underlying hematopoiesis, we employed DensityPath³⁴ to reconstruct the optimal cell state-transition trajectory. This method identified a geodesic minimum spanning tree (MST) of representative cell states (RCS) on the density landscape, establishing a least-action path characterized by minimal transition energy associated with cell fate decisions. Prior to applying DensityPath, we integrated four oligo-hashing reference samples using Seurat v5. Principal component analysis (PCA) was performed on the integrated dataset, and dimensions were selected by identifying the largest eigenvalue λ_i where the difference between consecutive eigenvalues fell below a threshold of 5 × 10⁻⁴. The resulting 25-dimensional PCA space was then used for non-linear dimensionality reduction via Elastic Embedding (EE)³³, a method that preserves both local and global data structures, to visualize the intrinsic structure of the scRNA-seq data in a 2-dimensional embedded space. In the EE-embedded space, DensityPath estimated the density surface and applied level-set clustering (LSC) to identify distinct high-density cell clusters, designated as RCSs¹²³. The cell state-transition trajectory was reconstructed by determining the MST of the RCS peak points, with geodesic distances computed on the density landscape. To resolve regions with low-density structures, such as the basophil lineage, which were obscured within the global structure, we extracted cells mapped to the GMP lineage branch and reapplied DensityPath to capture fine-scale structures. This process identified basophil-specific local clusters, which were then integrated with the global trajectory to yield a refined and comprehensive cell trajectory. Each cell was subsequently mapped to the nearest waypoint along the trajectory based on its geodesic distance. Using the fixed peak point of the HSC cluster as the starting cell, we computed pseudotime by calculating the geodesic distance between the projected positions of the starting cell and any given cell along the state-transition path. While LSC effectively extracted high-density RCSs as landmarks of the complex density landscape, and the identified RCSs provided an effective representation of subpopulations and gene expression, LSC did not fully capture cell type divisions. To address this, we employed mean-shift (MS) clustering^124,125, a density-based approach, to refine clustering. Instead of clustering cells directly, MS was applied to the waypoints along the trajectory, and cluster labels were then assigned to the corresponding cells mapped to those waypoints. For the basophil lineage, waypoints along the GMP branch were further refined via MS clustering to enhance clustering granularity. The final cell clustering results are referred to as density clusters. After obtaining cell cluster labels for the reference dataset, we used Symphony¹²⁶ to map additional scRNA-seq datasets onto the reference. This approach facilitated the transfer of MS-based density cluster labels and EE embedding to these datasets. Using the predicted EE embedding, cells were mapped onto the cell state-transition path, enabling pseudotime calculation based on their geodesic distances. In our R package HemaScribe, we provide a HemaScribe function to annotate cells and a HemaScape function to map query scRNA-seq datasets to our reference EE embedding, identify density clusters, and calculate pseudotime. This package is compatible with Seurat v5 objects and can be found at github.com/RabadanLab/HemaScribe along with the reference data.

Quantitative benchmarking of inferred trajectory trees against a lineage-tracing reference using random-walk fate propagation

To enable quantitative benchmarking of inferred trajectory graphs against a lineage-tracing reference, we employed a random-walk framework on directed graphs to quantify the fate structure implied by each trajectory. Each inferred trajectory was represented as a directed graph, with vertices corresponding to trajectory states and directed edges encoding state-to-state progression. Edge weights were given by the method-specific distances used during trajectory construction. For each vertex, we estimated empirical cell-type compositions from observed cell counts and normalized them to obtain vertex-level probability distributions over cell types. Fate propagation along each trajectory was modeled as a random walk on the directed graph. For inferred trajectories, transition probabilities from vertex i to each outgoing neighbor j were defined as proportional to exp(−βw_ij), where w_ij denotes the edge weight, and were row-normalized across all outgoing edges of vertex i. For the in vivo Hoxb5-tdTomato lineage-tracing reference³⁸, transition probabilities were instead constructed directly from the original edge flux/probability matrix, restricted to edges present in the tdTomato graph and row-normalized, without applying the exponential distance-based transformation. To accommodate potential cycles in the reference graph, terminal fates were defined as vertices belonging to terminal strongly connected components (SCCs), i.e., SCCs with no outgoing edges to other components. These vertices were treated as absorbing states. Under this formulation, which treats terminal states as absorbing, we analytically computed the probability of eventual absorption from every vertex into each absorbing (terminal-SCC) vertex using the standard fundamental-matrix solution for absorbing Markov chains. To obtain biologically interpretable fate outcomes, absorption probabilities into absorbing vertices were combined with the cell-type compositions of those vertices, yielding for each starting vertex a probability distribution over terminal cell types rather than over individual absorbing vertices. Vertex-level fate distributions were then aggregated to obtain conditional fate distributions at the level of starting cell types, P(final cell type | start type), by weighting vertices according to their composition of the given start type. Specifically, vertices with a higher fraction of a given starting cell type contributed proportionally more to that start type’s aggregate fate distribution, while vertices were otherwise equally weighted. To compare inferred trajectories with the tdTomato lineage-tracing reference, we computed the Jensen–Shannon (JS) divergence between the model-predicted and reference P(final cell type | start type) distributions for each starting cell type. These per–start-type divergences were aggregated into a single overall score by weighting each starting cell type according to its relative prevalence in the reference data, yielding a weighted overall JS divergence that summarizes agreement in global lineage fate structure. To assess uncertainty attributable to finite sampling of fate outcomes, conditional on the model-predicted distributions, we performed a parametric bootstrap at the level of P(final cell type | start type). For each starting cell type, fate counts were simulated via multinomial sampling from the corresponding model-predicted distribution, using a fixed total sample size allocated across starting cell types in proportion to their prevalences in the reference data. For each bootstrap replicate, sampled counts were converted to empirical fate distributions, and the weighted overall JS divergence was recomputed. Repeating this procedure yielded an empirical sampling distribution of the JS divergence, from which standard errors and confidence intervals were estimated. In addition, we constructed a null bootstrap distribution by sampling around the reference P(final cell type | start type) distributions, representing the expected variability in JS divergence when the inferred model and reference are identical and observed differences arise solely from sampling noise. One-sided p-values were computed as the probability that the null bootstrap divergence exceeded the observed divergence.

Validation of pseudotime inference using independent temporal references

To assess the accuracy of pseudotime inference, we compared pseudotime values against two orthogonal sources of temporal information: a computational potency score and direct measurements from lineage-tracing experiments. First, we mapped cells from external reference datasets onto each trajectory and computed their CytoTRACE2 scores³⁹, which estimate cellular developmental potency from transcriptomic data. We then calculated the negative correlation between the assigned pseudotime and CytoTRACE2 score for each method, where stronger negative correlations indicate better alignment of the pseudotemporal axis with the loss of developmental potential. Second, we evaluated pseudotime against real chronological time from two independent lineage-tracing studies. We utilized the in vivo Hoxb5-tdTomato lineage-tracing reference³⁸, which provides cells harvested at defined intervals post-label induction but lacks clone-level information. For each time-series dataset, we mapped tdTomato⁺ cells onto the trajectory, computed their pseudotime values, and evaluated the correlation between pseudotime and experimental time across all sampled time points. To account for sampling variability, we performed 1,000 iterations of subsampling 90% of cells per time point before mapping and correlation analysis, reporting the mean and standard deviation of the correlation coefficient. We also utilized the in vivo LARRY dataset⁴⁰, which contains clonal barcodes allowing the tracking of sister cells across time. We mapped all cells onto each trajectory, and for each clone present at multiple time points, we calculated the correlation between the pseudotime of its cells and their respective experimental sampling times. This provides a clone-resolved assessment of whether pseudotime progression reflects actual temporal dynamics during differentiation.

Identification of surface markers for HemaScape nodes

We downloaded a recently published CITE-seq dataset⁴⁸ from the Gene Expression Omnibus repository (accession numbers GSM8252385, GSM8252387, GSM8252389, GSM8252391, GSM8252393, GSM8252395, and GSM8252397). After merging, cells with anomalous (exceeding 5 MADs away from the median) nCount_RNA, nFeature_RNA, or nCount_ADT, as well as cells with mitochondrial gene percentage more than 5 MADs above the median or greater than 10%, whichever threshold was lower, were excluded for quality control. The RNA-seq data was normalized using Seurat::NormalizeData with default parameters; the ADT data was normalized with Seurat::NormalizeData with normalization.method=”CLR” and margin=2. We then ran HemaScribe and HemaScape on the RNA-seq data to assign each cell to a HemaScape node. We used the Seurat::FindMarkers function to identify surface markers that were differentially expressed in the ADT data contrasting nodes 1+2 versus node 9. To evaluate cell enrichment using novel CD24 and CD62L candidate surface markers, we sorted 8,000 cells of each type into 1.5 ml tubes, added 0.5 μg of different Total-SeqB antibodies (BioLegend, B0301 for LSK, 155831; B0302 for CD24⁺ LSK, 155832; B0303 for CD62L⁺ LSK, 155833) and then rested them on ice for 1 hour. After this, we washed cells three times with cell staining buffer and pooled them together into a single tube. From this, we partitioned cells, prepared libraries, and sequenced samples as described above for scRNA-seq samples. After obtaining fastq files, we ran the Cellranger count pipeline (version 9), used the same quality control pipeline as outlined above, demultiplexed cells using MULTI-Seq as described above, and ran HemaScribe and HemaScape to map cells to cell types and density cluster nodes.

PAGA trajectory and connectivity changes with different perturbations

To analyze the connectivity among the MS-based density clusters, we employed Partition-based Graph Abstraction (PAGA)³⁵, a graph-based method that abstracts high-dimensional manifolds inherent in scRNA-seq data and quantifies connectivity between density clusters. For oligo-hashing reference samples, we performed PAGA using the first 50 principal components (PCs) and density clusters after data integration. To align with the DensityPath framework, we visualized the PAGA graph backbone in the embedded EE space, where edge weights represent the PAGA connectivity between cell groups. To assess the effects of different perturbations, we applied PAGA analysis to paired treated-control samples. For each sample, we used the first 50 PCs derived from the scRNA-seq data and density clusters predicted by Symphony. Connectivity changes were evaluated by comparing the edge weights between treated and control conditions, enabling us to quantify the perturbation effects on cell differentiation.

Evaluation of differentiation tendency

To evaluate transcriptomic bias towards terminal differentiation, we used Palantir v1.4.0 as described by the authors⁴⁶. For each dataset (Table S2D), we used the EE embedding created by Symphony mapping onto the HemaScape reference dataset, and we defined cells representing start and end points by manual selection. Possible terminal points were therefore defined in the GP, cMoP, erythroid, megakaryocyte, CLP, and basophil branches, and we scaled Palantir entropy for cells in each dataset for presentation. To estimate quantitative changes in hematopoietic differentiation, we created random walk models based on a simplified tree structure containing 15 nodes. For each node, we estimated proliferation rates using a transcriptomic score of cell cycle activity. To do so, we followed the tutorial from the R package UCell: we imported count matrices from 6 test scRNA-seq datasets (representing 3 WT steady state LK and 3 LSK datasets) into SingleCellExperiment objects and scored each cell with the KEGG Cell Cycle geneset (Table S2B) using ScoreSignatures_UCell function. Using our HemaScribe annotation for the same datasets, we then calculated the mean cell cycle score for each cell type. In parallel, we measured the cell cycle status of the same HSPC cell type by flow cytometry in Fucci2 mice⁴⁵, identifying actively cycling cells by their expression of Geminin-mVenus. Next, we explored the relationship between cell cycle score and Fucci2 data by fitting linear, logarithmic, or quadratic functions to these data and calculating the Akaike information criterion (AIC) for each. The logarithmic function defined by y=0.0172.log(x) + 0.0298 had the optimal AIC (data not shown), and we selected this relationship for use with all datasets. For cells assigned to each node in each dataset, we then interpolated the cell cycle activity using the mean cell cycle score for each node using this function. Values greater than 1 (representing 100% of cells engaged in the cell cycle) were assigned a value of 1. To estimate cell death rate in each node, we first estimated cell death rate in each HSPC cell type by ApoTracker Green flow cytometry in control mice and mice subjected to each studied perturbation. Next, we calculated an average cell death rate for each node based on its HemaScribe cell type composition and the cell death rates measured by flow cytometry. For this purpose, GP, cMoP, mGMP, and basophils identified by HemaScribe were all treated as GMP, whereas immature B cells were considered as CLP. To estimate the likelihood of cell distribution from an upstream node to its downstream nodes, we used the connectivity matrices generated for each dataset by PAGA analysis, as outlined above. For this purpose, we set all values below the diagonal to 0 (enforcing unidirectional flow), we filtered out connections that did not appear in our tree structure, and we normalized connectivity for each node so that all its downstream connections summed to 1. To estimate differentiation rate for each node, representing the likelihood of cells proceeding from one node to any downstream node(s), we implemented an iterative stochastic simulation approach, where we calculated the size of each node in each dataset using stochastic samples of the following parameters driven by binomial processes: proliferation rate, cell death rate, differentiation rate (set to an arbitrary value of 0.3 in the first iteration), and cell distribution using the normalized connectivity matrix. The starting node size for each iteration was derived from the actual size of each node in LK scRNA-seq datasets, which were taken to represent the natural frequency of HSPCs. Cells transitioning from an upstream node were included in the size calculation for their downstream node(s). After each iteration, the error in node size was calculated by comparing final node size with the actual node size, and the differentiation rate was then adjusted according to the direction of the error. Each simulation was run for 10,000 iterations across 10 runs, and differentiation rates were averaged for the final 50 iterations for each run to generate final values for each node. Node parameters used for simulations are provided in Table S2C (steady state) and Table S2F-J (perturbations). To evaluate differentiation tendency over the time course of G-CSF and pIC treatment using optimal transport, LK scRNA-seq dataset for each perturbation were first normalized with SCTransform in Seurat v5 and then integrated using the PrepSCTIntegration, FindIntegrationAnchors, and IntegrateData functions. Matrices were subset to 2,000 highly variable genes that were selected with the SelectIntegrationFeatures function. These data were then used as input for Waddington-OT⁵⁶, supplying metadata on time of sampling after starting treatment and HemaScape node identity for each cell. Proliferation and cell death rates were estimated using provided gene scores and scoring methods. Transport maps were calculated with standard settings (epsilon = 0.05, lamba1 = 1, and lambda2 = 50) and with 20 growth iterations. Fate matrices and transition tables were generated by selecting day 0 cells and extracting predicted fates at day 4.

Evaluation of transcriptional response

To understand which HemaScape nodes were mounting the greatest transcriptional responses to perturbation, we used the function calculate_auc in the R package Augur⁵⁵, entering the perturbation as label_col and the HemaScape node identity as cell_type_col.

Inference of transcription factor activity

To estimate transcription factor activity in transcriptomic data, we performed gene regulatory network inference using SCENIC via the fast Python implementation pySCENIC^41,129. Starting with the scRNA-seq expression matrices converted to loom format, we ran the commands pyscenic grn, pyscenic ctx --mask_dropouts, and pyscenic aucell with the auxiliary mm10 and hg38 genome annotations and v10 motif datasets retrieved from https://resources.aertslab.org/cistarget/. To estimate the effects of removing critical transcription factor nodes, we used CellOracle version 0.18.0^[77] implemented via Docker image and using the base gene regulatory network generated by the authors from HSPCs¹²⁸. For each dataset, we used the EE embedding generated by Symphony mapping onto the HemaScape reference dataset, as outlined above. We performed in silico perturbations as outlined in the tutorial, and we visualized the results of single perturbations using a digitized grid, for which scale and min_mass values are shown for each dataset in Table S2D. To evaluate the relative importance of different transcription factors in myelopoiesis in different perturbation conditions, we performed an in-silico screen for a candidate list of 33 transcription factors selected for their known importance in myelopoiesis (Table S3F). For each dataset, we evaluated the impact of in silico perturbations relative to developmental flow defined by calculation of a pseudotime gradient starting in the HSC cluster, with selected starting cells shown in Table S2D. For the in silico screens, we then defined a myelopoiesis trajectory that included the following density clusters: 1, 2, 3, 9, 10, 11, 12. After running the screens, we extracted the negative perturbation score p-values for each transcription factor (representing the significance of disruption relative to developmental flow in the myelopoiesis trajectory) and scaled them between 0 and 1 for each perturbation.

Analysis of epigenetic poising

To test whether cells in any HemaScape node showed poised chromatin regions for response to perturbation, we first used the FindMarkers function in Seurat to derive lists of upregulated genes for all HSPCs together in perturbation datasets compared to their matched controls, excluding any cells labeled as “NotHSPC” in the HemaScribe fine classifier. Next, we used the function promoterRegions from the R package Rsubread v2.22.1 to derive promoter regions for the murine mm10 genome for the top 500 upregulated genes, which we defined as regions extending 2000 bp upstream and 200 bp downstream from the transcription start site. We also obtained promoter regions for 500 randomly selected genes to act as a measure of background accessibility. Next, we reanalyzed a published single-cell LK/LSK multiome dataset²⁸. Samples for steady state adult LK and LSK cells were analyzed in Signac version 1.14.0^[129] with the following QC parameters: ATAC count >1000 and <100,000 per cell, nucleosome signal <2, TSS enrichment >1. Samples were integrated using the IntegrateEmbeddings function, and peaks were called separately for each dataset with MACS2 then combined to create a merged peak file, which was used to count peaks in the Signac object. After integration, principal component analysis was performed to identify 30 components, followed by UMAP using components 2 through 30. The RNA assay was processed as described above for scRNA-seq datasets and used to annotate cell types with HemaScribe. We first extracted all peaks from the Signac object and intersected these regions with the promoter regions for genes upregulated in each perturbation using the intersect function in bedtools v2.31.1, with the options -wa -a, to retain Signac peaks that overlapped with promoter regions. We then scored the accessibility of the overlapping promoters in the Signac object using the AddChromatinModule function in Signac. We extracted these scores for each HemaScape node and calculated the mean difference in the accessibility score for each perturbation compared to the sample of promoters from 500 random genes to correct for differences in background accessibility across cell types.

Non-negative matrix factorization in mouse and human datasets

To uncover possible shared molecular signatures in emergency myelopoiesis, we first integrated 25 separate scRNA-seq datasets, comprising LK and LSK datasets from 9 different perturbation conditions (Table S3A). After applying quality control procedures outlined above, these datasets were transformed with SCTransform in Seurat v5 and then integrated using the PrepSCTIntegration, FindIntegrationAnchors, and IntegrateData functions. We created a shared list of highly variable genes across all these datasets by also running the SelectIntegrationFeatures function with a liberal threshold of 10,000 genes, from which 7,184 genes were retained. From the integrated dataset, we then extracted the SCT-normalized layer that was subsetted against the list of highly variable genes and performed non-negative matrix factorization (NMF) using the nmf function from the R package RcppML v0.5.6^[130]. The optimal number of factors was selected by running cross-validation using the crossValidate function. The h matrix was added to the integrated Seurat object as metadata, whereas the w matrix was added as a dimensionality reduction. Each NMF factor was characterized by performing the following procedures: (1) assessment of variance according to HemaScribe cell type and perturbation by calculation of the Kruskal-Wallis test statistic using the kruskal.test function from the stats R package v4.3.1, (2) evaluation of enrichment for Gene Ontology, KEGG, Reactome, and Hallmark pathways in the top 500 genes loading on each factor, using the GSEA function from the R package clusterProfiler v4.8.3^[131,132], with the scoreType option as “pos” to account for the absence of negative values, (3) evaluation of cytokine dictionary enrichment by Immune Response Enrichment Analysis (IREA)⁶⁷ (www.immune-dictionary.org) in its interactive interface, using the “macrophage” cell type, and (4) Pearson correlations between h matrix scores and SCENIC transcription factor regulon activity for each factor in the same integrated dataset using the cor function in stats. To understand how NMF factors changed in activity with myeloid differentiation, we also calculated the most active factors based on h matrix scores across the pseudotime gradient created with the EE embedding as described above, which were presented as stream plots. To determine the extent of evolutionary conservation of top loading genes in the murine NMF model, we used the getBM function in biomaRt v2.65.0 to query the Ensembl database of species orthologs for the top 100 loading genes for M12, M18, M24, and M40. We extracted information for Homo sapiens, Danio rerio, Xenopus tropicalis, Gallus gallus, Rattus norvegicus, Astyanax mexicanus, Poecilia formosa, Tetraodon nigroviridis, Takifugu rubripes, Gadus morhua, Pan troglodytes, Tursiops truncatus, Notamacropus eugenii, Podarcis muralis, Latimeria chlamunae, and Petromyzon marinus, and we then determined if there was at least 1 annotated ortholog for each of the murine genes. We performed the same analysis with 100 randomly selected genes to represent the background level of conservation and, for presentation purposes, we obtained a dendrogram of evolutionary relationships of vertebrate species from FigShare (https://figshare.com/articles/dataset/Vertebrates_species_trees/22736555). The same procedures were followed for creation of a NMF model using human perturbation datasets (Table S3H). The correspondence of human and murine NMF factors was evaluated by converting mouse gene names to human orthologs as above and then calculating Spearman correlations between factor gene loadings in the w matrices of both models (Table S3J). To project our mouse NMF model onto independent query datasets (Table S3D), we intersected genes present in the query dataset and NMF model and then performed non-negative least squares regression using the w matrix and the query expression matrix to reproduce a pseudo-h matrix for the query dataset. This was implemented with the nnls function in the R package nnls (version 1.6). Scores for different perturbations and cell types were compared from these h matrices using Wilcoxon’s signed rank test or Kruskal-Wallis test with post hoc Dunn’s test and Bonferroni correction, all implemented with the stats R package version 4.3.1.

Consensus non-negative matrix factorization

The same input data were also used to perform consensus (c)NMF using the cNMF package according to the author’s instructions⁶⁸, with k=40 factors showing the best balance between stability and error. Spearman correlations were calculated between factor gene loadings in the w matrix in the original model (M1–40) against factors in the cNMF model (C1–40) using the R stats package, and factors were paired with their best match using the Hungarian method implemented by the solve_LSAP function in the R package clue (version 0.3–66).

Generation of signatures and human orthologs from mouse NMF factors

To generate gene signatures from NMF factors, we first intersected the top loading genes in selected NMF factors according to scores in the w matrix against the most significantly upregulated genes in target cell types in relevant perturbations, taking the top 50 genes that appeared in both lists according to loading score and adjusted p-value. Thus, for M12, we compared 20-day IL1 to control; for M18, we compared 7-day IL1, 16 hours LPS, and pIC day 3 to control; for M40, we compared 16 hours LPS and pIC day 3 to control; and for M24, we compared 5FU day 8, G-CSF day 2, G-CSF day 4, 16 hours LPS, and pIC day 3 to control. We mapped mouse gene symbols to human gene symbols using the getLDS function from the R package biomaRt v2.65.0, and we manually curated these results to search for any missing orthologs. A complete list of signatures and coefficients is shown in Table S3G.

Gene set enrichment analysis in bulk RNA-seq samples

To test enrichment of mouse NMF factor genes or human ortholog signatures in bulk RNA-seq datasets, we used the GSEA function in clusterProfiler version 4.8.3, and we generated plots using enrichplot version 1.20.3. Published RNA-seq datasets were downloaded from the Gene Expression Omnibus (GEO) (Table S3E). Where possible, author-generated lists of differentially expressed genes were used as input. Where this was unavailable, counts tables were downloaded and re-analyzed using DESeq2 v1.40.2. If counts were unavailable, fastq files were downloaded, trimmed with trim-galore, and pseudo-aligned to the mouse mm10 transcript reference using Salmon, before differentially expressed genes were identified with DESeq2. If only normalized counts were available, voom, lmFit, and eBayes were used in the R package limma version 3.58.1 to perform differential expression testing.

Analysis of published cross-linking and immunoprecipitation (CLIP)-sequencing for YBX1

Fastq files were downloaded from GSE63604 for n=3 biological replicates of CLIP-seq of human MDA-231 cells⁷⁹ and trimmed with trim-galore. Reads were aligned to the human hg38 genome using Bowtie-2 (version 2.5.4), with settings –local, --very-sensitive-local, -k 1, --no-unal, -p 8. Blacklisted regions were removed and bigwig files were created from bam files using the bamCoverage function from deepTools (version 3.5.6), and the R package trackplot (version 1.6) was used to generate genomic track plots. Enrichment for YBX1 binding sites in M24 genes was evaluated by converting the top 100 loading genes of M24 to 91 human orthologs, and then creating a contingency table against genes that had a YBX1 peak called in the original publication (available in GEO under GSE63604). Enrichment was tested with Fisher’s exact test in the R package stats.

Prognostic evaluation of myelopoiesis signatures in human disease

We conducted survival analyses examining the impact and prognostic value of the emergency myelopoiesis signatures derived from our NMF analysis above in two cohorts of human AML patients. Data from 136 samples in the TCGA-LAML project⁸⁹ and 273 BM aspirates at time of diagnosis from Beat AML^90,91 with matched bulk RNA transcriptomic data and clinical survival data were retrieved from cBioPortal^133,134 (TCGA-LAML: cBioPortal version 6.0.12). We normalized the NMF coefficients for genes in each signature to sum to unity. Then, each sample was given a score defined as the weighted average of the expression levels (in log(1+TPM) units) of the genes in the signature using the normalized NMF coefficients as weights. To analyze the relationship between these scores and other variables, we generated oncoplots using oncoPrint from ComplexHeatmap¹³⁵ to visualize whether our signatures are correlated with other explanatory features such as age, sex, or mutation status, and performed two-sided Wilcoxon rank sum tests to quantify their associations. For analysis of AML prognosis, we first performed univariate Cox proportional hazards regression with covariates for overall survival in months on the Beat AML cohort to identify other potential mutational correlates of overall survival among single nucleotide polymorphisms that occurred in at least 5% of the samples using the surv_fit function from the survival R package^136,137; p-values were adjusted for false discovery rates using the Benjamini-Hochberg procedure in stats::p.adjust. We obtained the Cox regression results on hSig24 alone and also for a multivariate analysis on hSig24 together with the mutational correlates that were found significant in the previous step. Age, sex, and prior history of malignancy or treatment were included as covariates in all of the Cox regression models, where available. We plotted Kaplan-Meier survival curves for the top 20% and bottom 20% for each signature respectively and conducted a log rank test for significance using the ggsurvplot function from survminer¹³⁸. To isolate the role of hSigs from TP53 mutation status, we selected TP53 wild-type (WT) patients from the TCGA-LAML and Beat AML cohorts. For these sub-cohorts, we repeated our analysis above by incorporating our hSig scores into individual Cox regression models for overall survival, derived hazard ratios and p-values, and plotted Kaplan-Meier survival curves for hSig24 in the TCGA-LAML and Beat AML sub-cohorts separately. In order to benchmark these results, we obtained ELN2022 risk classifications for patients in the two cohorts from previously published sources^139,140 and trained additional Cox regression models using both the ELN2022 risk, clinical covariates, and each signature score on the full cohorts. We compared their performances to a baseline model trained using only the ELN2022 risk and covariates using likelihood ratio tests and ANOVA. To compare the prognostic value of our scores to that of existing risk scores, we repeated this benchmarking analysis with other gene expression-based AML prognosis scores previously described in the literature^98–101. Moreover, we expanded our analysis and performed similar survival analyses for other non-myeloid cancers represented in TCGA, to serve as negative controls^141,142. We also conducted survival analyses for pediatric myeloid malignancies including 1786 cases of pediatric AML from the TARGET project^143,144. These secondary cohorts were retrieved from the cBioPortal and NCI Genomic Data Commons^145,146. Finally, we also evaluated these myelopoiesis signature scores on bulk RNA-seq data from other previously reported cohorts totaling 123 adult AML patients at diagnosis without matched survival information and stratified the score distributions using a previous molecular AML classification scheme based on cell type deconvolution⁸⁸.

Quantification and statistical analysis

Data are represented as means ± standard deviations (S.D.) or standard error of the mean (S.E.M.), or as violin plots with the center line representing the median, using R studio for molecular data or GraphPad Prism (version 10.4.2) for all other data. Circles on bar graphs represent biological replicates. For experimental data, Student’s t-test was used when 2 groups were compared. Either one-way ANOVA with Tukey’s post-hoc test, or Kruskal-Wallis test with Dunn’s post-hoc test were used to compare 3 or more groups. Data collection and analysis were not performed blind to the conditions of the experiments. Approximate sample size was predetermined for most experiments based on our experience of mouse EM perturbations, where a standard deviation of less than 10% can be expected for most measurements. To show a difference of at least 10%, with α = 0.05 and 80% power, 6–11 independent biological replicates are required per group.

Supplementary Material

2

Table S1. Annotation approach in HemaScribe, related to Figures 1 and 2.

• Tab A: List of bulk RNA sequencing samples used for broad classifier in HemaScribe.

• Tab B: List of single-cell RNA sequencing datasets annotated in this study.

• Tab C: Cell type labels for all steady state and perturbation datasets generated or analyzed in this study.

• Tab D: Genes used for hematopoietic score in HemaScribe.

3

Table S2. Creation of HemaScape tree and analysis of perturbation effects, related to Figure 3 and 4.

• Tab A: PAGA connectivity matrix for HemaScape tree, incorporating 3 LSK and 3 LK datasets.

• Tab B: Genes in the KEGG Cell Cycle geneset used for scoring single cell RNA sequencing datasets.

• Tab C: Node parameters for random walk models in steady state datasets.

• Tab D: Cells used as start and termination points for Palantir entropy analysis and start points for CellOracle analysis.

• Tab E: Perturbation conditions analyzed in this study.

• Tab F: 5FU day 8 versus control.

• Tab G: IL1 day 20 versus control.

• Tab H: LPS 16 hours versus control.

• Tab I: pIC days 2–4 versus control.

• Tab J: G-CSF days 2–4 versus control.

4

Table S3. Datasets related to generation of NMF models, related to Figures 5, 6, and 7.

• Tab A: Datasets used for mouse NMF model.

• Tab B: Loading values for genes on each NMF factor in model.

• Tab C: Description of NMF factor classification and major molecular processes.

• Tab D: Datasets onto which murine NMF model was projected.

• Tab E: Published RNA sequencing datasets analyzed in this study for factor or signature enrichment.

• Tab F: Transcription factors used for in silico screen using CellOracle.

• Tab G: Human emergency myelopoiesis gene signatures.

• Tab H: Datasets used for human NMF model.

• Tab I: Loading scores for NMF factors for human NMF model.

• Tab J: Spearman correlation coefficients comparing gene loading values for NMF factors from murine (M) and human (H) models.

5

Table S4. Clinical characteristics and signature scores for TCGA-LAML and Beat AML cohorts, related to Figure 7.

• Tab A: TCGA-LAML

• Tab B: Beat AML

Document S1. Figures S1-S16 with figure legends.

HIGHLIGHTS.

• New HemaScribe method for murine HSPC annotation in scRNA-seq datasets

• Different emergency myelopoiesis (EM) inducers target different HSPC populations

• Unique and shared transcriptional modules are enacted by different EM inducers

• A myeloid progenitor EM module informs outcome in human acute myeloid leukemia

RESOURCE AVAILABILITY

All newly generated datasets are deposited in the Gene Expression Omnibus (GEO) under GSE296404 and GSE298048. Published datasets are provided in the Key Resources Table. Reference data are provided in an interactive website at http://3.233.5.190:8050. Functions for HemaScribe annotation of cells in scRNA-seq datasets and mapping to the HemaScape differentiation landscape are available in the R package “HemaScribe” (github.com/RabadanLab/HemaScribe). Correspondence and requests for materials should be addressed to E.P. (ep2828@columbia.cumc.edu).