To create the AML scAtlas, we integrated published scRNA-seq data of primary AML bone marrow samples, from 16 suitable high-quality studies (see Materials and methods), comprising 159 AML samples (Figure 1A; Source data 1). Where on-treatment time points were available, we selected only diagnostic samples to establish a reference atlas of primary AML at diagnosis. If studies had healthy donor bone marrow samples, these were included, alongside data from healthy bone marrow samples from four additional scRNA-seq studies (Source data 1) to enable comparisons between malignant and healthy bone marrow populations. After cell filtering and quality control, the AML scAtlas contains data from 748,679 high quality cells derived from a total of 20 different scRNA-seq studies (Velten et al., 2021; van Galen et al., 2019; Beneyto-Calabuig et al., 2023; Stetson et al., 2021; Zheng et al., 2017; Petti et al., 2019; Jiang et al., 2020; Johnston et al., 2020; Pei et al., 2020; Li et al., 2023; Lasry et al., 2023; Fiskus et al., 2023; Naldini et al., 2023; Mumme et al., 2023; Zhang et al., 2023; Li et al., 2022; Oetjen et al., 2018; Setty et al., 2019; Caron et al., 2020; Figure 1—figure supplement 1A). Each sample was assigned to an AML clinical subtype, based on the recent European Leukemia Net (ELN) clinical guidelines (Döhner et al., 2022), and classified into the corresponding prognostic risk group. This resource captures a broad range of molecular subtypes of AML and spans different age groups, including both pediatric and adult AML cases (Figure 1B–C). Overall, this is the largest dataset to date for exploring AML biology at single-cell resolution.

In the initial analysis of the combined dataset, batch effects were noted with study-specific clustering, which was quantified using several benchmarking metrics (Figure 1—source data 2; Figure 1—figure supplement 1A and B). Even within samples of the same study, sample-wise clustering was noted (Figure 1—figure supplement 1C). To address this, we benchmarked several widely used batch correction methods (Figure 1—source data 2; Figure 1—figure supplement 2A–C), identifying scVI as the best method for this dataset (Figure 1—source data 2). We, therefore, employed scVI to correct for batch effects, before clustering and cell type annotation in the AML scAtlas, by using the consensus of multiple annotation tool results (Figure 1—source data 2), verified using cluster-wise marker gene expression (Figure 1D–E).

Cell type proportions analyses across the clinically relevant subtypes in the dataset show that the AML subtypes were significantly biased towards myeloid cell types (CMP, MEP, GMP, ProMono, CD14+ Mono, CD16+ Mono, cDC, Erythroid) with each subtype exhibiting a clear predominant cell type consistent with AML clonal expansion (Figure 2A; Figure 2—figure supplement 1A–B). In contrast, healthy donor samples had more balanced lineage proportions, with lymphoid cells (T, B, NK, ProB, pDC, Plasma) well represented (Figure 2A; Figure 2—figure supplement 1A–B). Given the established critical role of HSPCs and LSCs in propagating AML, and their importance as therapeutic targets (Montefiori et al., 2021), we focused on HSPC clusters for further analysis (Figure 2B). To identify LSCs, we applied a curated reference profile of leukaemic stem and progenitor cells (LSPCs) (Zeng et al., 2022; Figure 2B–C) and correlated this with calculated LSC6 (Elsayed et al., 2020) and LSC17 (Ng et al., 2016) scores for each cell (Figure 2D). We then compared the proportions of HSPC/LSPCs across different AML subtypes and risk groups, as defined by the ELN clinical guidelines (Döhner et al., 2022; Figure 2E). Higher-risk subtypes displayed a higher proportion of LSCs compared to favorable risk disease (Figure 2E–F).

The AML scAtlas enables robust comparison of adult and pediatric AML. We hypothesized that in adolescents and young adults with t(8;21) AML, the potential for either in-utero or postnatal HSPC origin disease might affect disease biology and prognosis. Thus, we sought to explore biological differences between pediatric and adult cases of t(8;21) AML, aiming to explain and potentially improve prognostication in adolescents and young adults. We selected samples with t(8;21) AML from the AML scAtlas, resulting in 105,663 cells from 13 adult cases (aged 20–67), seven adolescent cases (aged 12–17), and three pediatric cases (aged 6–8) (Figure 3A–C). Where gender information was not available, this was inferred from ChrY/XIST gene expression (Figure 3—figure supplement 1A). Several adult samples underwent CD34 selection in original studies, excluding more differentiated cell types (mature lymphoid populations, monocytes, granulocytes) in these samples. Thus, these cell types were excluded from comparative analysis, focusing only on HSPCs and myeloid progenitors (CMP, GMP, MEP), which were well represented in all studies (Figure 3C).

We reconstructed the GRN for the t(8;21) subset using the pySCENIC (Van de Sande et al., 2020) pipeline, which is a Python-based efficient implementation of original SCENIC method (Aibar et al., 2017). It is a state-of-the-art method for network inference from scRNA data, popular in the community (Hamed et al., 2022; Barnett et al., 2024; Zhang et al., 2024) and has shown strong performance in a recent benchmarking study (Nguyen et al., 2021). SCENIC’s three major steps are: First, it identifies groups of co-expressed genes as potential targets of a transcription factor (TF). Second, it refines these groups of genes based on the enrichment of the corresponding TF binding motif, forming ‘regulons.’ Third, it uses the AUCell method (embedded within SCENIC) to quantify the activity of each regulon in every cell. AUCell calculates the Area Under the Curve (AUC) for the regulon’s genes set in a ranking of all genes by expression for each cell. The top 20 regulons for each age groups were selected based on the regulon specificity score (RSS) (Figure 3—figure supplement 1B). Unsupervised clustering on the Zscore normalized regulon activity score matrix revealed clear differences in the GRN across different age groups (Figure 3—figure supplement 1C). We hypothesize that the differences in the GRN might reflect differences in the pre- or postnatal developmental origins of the disease. Additional testing of GRN inference from individual studies shows that the high number of cells refines the overall GRN (see Methods; Figure 3—figure supplement 1D–E).

To define gene regulatory programs (co-occurring gene modules, defined by a transcription factor and its targets) which are specific to different age groups (termed ‘regulon signature’), we used the clustered dendrogram to select the regulon clusters most associated with the pediatric (below 10 years-old) and adult samples (over 18 years-old) (Figure 3D; Figure 3—figure supplement 1C). The pediatric regulon signature, proposed to represent in-utero origin t(8;21) AML (henceforth termed ‘inferred-prenatal’), includes 16 regulons defined by a distinct group of hematopoietic transcription factors (TFs) (TRIM28, CTCF, RAD21, SOX4, TAL1, MYB, FOXN3, JUND, BCLAF1, ZBTB7A, IKZF1, MAZ, REST, YY1, CUX1, KDM5A), many of which have clearly defined roles in HSPCs and AML (Lu et al., 2018; Fisher et al., 2017; Kumar et al., 2023). The adult regulon signature, presumed representative of the postnatally acquired t(8;21) AML (henceforth termed ‘inferred-postnatal’), combines three discrete clusters of regulons (YBX1, ENO1, and HDAC2; GATA1, POLE3, TFDP1, MYBL2, E2F4, and KLF1; IRF1, STAT1, IRF7, MAFF, ATF4, TAGLN2, SPI1, and KLF2), defined by TFs previously implicated in various hematopoietic, leukemic and inflammatory processes (Ning et al., 2011; Fischer et al., 2019; Fang et al., 2018). Importantly, both signatures contain key components of the AP-1 complex, which is heavily implicated in the biology of t(8;21) AML (Ptasinska et al., 2019; Martinez-Soria et al., 2018) and undergoes dynamic changes during aging (Patrick et al., 2024). Samples of 6 individuals aged 12–17 clustered with the pediatric samples and showed enrichment for the inferred-prenatal signature (Figure 3D), suggesting that older adolescents (up to age 17 in our cohort) more closely resemble pediatric AML with t(8;21) and remain biologically distinct from adult-onset disease. This implies that the inferred in-utero origin of t(8;21) AML can also be present in AML diagnosed in older children.

We next sought to externally validate our age-associated regulon signatures in a larger cohort of patients. Bulk RNA-seq samples were obtained from the TARGET (Bolouri et al., 2018) and BeatAML (Burd et al., 2020) cohorts, selecting bone marrow samples taken at diagnosis in line with AML scAtlas data (n=83; Source data 2). We applied the AUCell algorithm from pySCENIC (Van de Sande et al., 2020) to calculate the activity of our pediatric inferred-prenatal and adult inferred-postnatal regulons in each sample. Unsupervised clustering of the bulk RNA-seq AUCell results revealed discrete clusters of samples that were highly enriched for our inferred-prenatal and inferred-postnatal origin-associated regulons (Figure 4A).

Given the limitations of most scRNA-seq platforms in detecting lowly expressed genes, notably TFs, we leveraged bulk RNA-seq samples to refine our identified gene regulatory networks by detecting differentially expressed regulon-associated TFs. We used our inferred-prenatal and inferred-postnatal signature clusters and performed differential gene expression analysis between these samples, using two widely used tools (DESeq2 Robinson et al., 2010 and edgeR Love et al., 2014) to ensure robustness of the results (Figure 4B; Figure 4—figure supplement 1A). We then compared differentially expressed regulon-associated TFs between the two groups and intersected this with the differential genes detected by each method. Although changes in TF expression are subtle (Figure 4B; Figure 4—figure supplement 1A), we identify significantly differentially expressed TFs which reflect the observed differences in regulon activity and indicate the most critical regulons in our age-related GRN signatures (Figure 4A–B; Figure 4—figure supplement 1B). This further delineated the inferred-prenatal signature to five key TFs (KDM5A, REST, BCLAF1, YY1, and RAD21), and the inferred-postnatal signature to eight TFs (ENO1, TFDP1, MYBL2, TAGLN2, KLF2, IRF7, SPI1, and YBX1).

We next performed gene set enrichment analysis (GSEA) on significantly differentially expressed genes as determined by edgeR (Love et al., 2014), to investigate pathways enriched in the inferred-prenatal samples compared to the inferred-postnatal ones (Figure 4C). Notably, inferred-prenatal samples showed increased expression of stemness-associated genes, and SMARCA2 target genes, a key player in HSC gene expression regulation and chromatin remodeling (Holmfeldt et al., 2016). SMARCA2 is also known to be upregulated during the fetal-to-adult HSC transition (Chen et al., 2019), implying that the observed SMARCA2 enrichment may indeed reflect the inferred fetal HSC cell-of-origin. Genes impacted by YY1 depletion were also downregulated compared to the samples of inferred-postnatal leukemia origin, which supports the identification of YY1 as an inferred-prenatal regulon (Figure 4C). To explore therapeutic implications, we performed GSEA using drug response signatures from published studies (Unnikrishnan et al., 2017; Williams et al., 2020; Xu et al., 2019; Zhang et al., 2020; Figure 4D). This analysis revealed that inferred-prenatal origin t(8;21) AML is enriched for genes associated with increased chemosensitivity to cytarabine, venetoclax, and daunorubicin.

We hypothesized that the increase in stemness-associated genes in the leukemia with the inferred-prenatal origin could be reflective of potential differences in the leukemic cell-of-origin and its impact on myeloid differentiation. We therefore performed cell type deconvolution using AutoGeneS (Aliee and Theis, 2021), with a curated LSPC reference profile (Zeng et al., 2022), to compare the cellular heterogeneity between prenatal and postnatal origin bulk RNA-seq samples (Figure 4E). This revealed a higher proportion of HSPC cell types (HSC, Prog), with a reduction in some differentiated myeloid cell types (ProMono-like, cDC-like) in the samples of inferred-prenatal origin (Figure 4E). To corroborate this finding, we examined cell type proportions in the original t(8;21) subset of AML scAtlas, confirming that cells with the inferred-prenatal signature comprise more HSCs than inferred-postnatal signature cells (Figure 4—figure supplement 1D–E). However, comparison of cell type proportions in this dataset is confounded by differences in sample processing as some studies performed CD34 selection, hence, there is more cell type diversity observed in the pediatric samples (Figure 4—figure supplement 1D–E).

We next used the scRNA-seq and scATAC-seq data from a recent cohort of pediatric t(8;21) AML patients (Lambo et al., 2023) at multiple clinical time points to uncover the enhancer-driven GRN (eGRN) in inferred-prenatal and inferred-postnatal origin t(8;21) AML (Source data 2). Initially, we identified two representative samples of our inferred-prenatal and inferred-postnatal signatures by using pySCENIC AUCell (Van de Sande et al., 2020) to measure the activity of our previously defined regulons. Unsupervised clustering of the AUC scores was used to infer whether each sample matched the regulon signatures, identifying one inferred-prenatal sample and one inferred-postnatal sample for downstream analysis (Figure 5—figure supplement 1A).

We then applied SCENIC+ (Bravo González-Blas et al., 2023), which integrates scRNA-seq and scATAC-seq to identify candidate enhancer regions and TF-binding motifs, linking TFs to target genes and identified enhancers. This creates enhancer-driven regulons (eRegulons), forming an eGRN. We applied SCENIC+ (Bravo González-Blas et al., 2023) to the leukemia samples with the inferred-prenatal and inferred-postnatal origin at diagnosis and relapse, keeping only regulons that showed a correlation between both modalities to retain only the most robust regulons (Figure 5—figure supplement 1B). This revealed several eRegulons across both patients (Figure 5—figure supplement 1D), many of which were patient-specific, particularly when comparing HSPC populations (Figure 5A). The inferred-prenatal sample displayed a specific HSC eRegulon profile. In contrast, the inferred-postnatal sample more closely resembled the corresponding Granulocyte-Monocyte Progenitor (GMP) (Figure 5A). Interestingly, at relapse the inferred-prenatal origin patient undergoes a chemotherapy-driven lineage switch to a lymphoid phenotype, which may suggest that the leukemia originated from a less committed progenitor (Figure 5A).

To identify clusters of closely related eRegulons, we computed the correlations between eRegulons enrichment. We identified two main clusters of eRegulons which correspond to different inferred-signature samples (Figure 5B; Figure 5—figure supplement 1C). For each eRegulon cluster, we used the associated target genes as input for gene ontology over representation analysis (ORA), to assess functional differences in the eGRN. This revealed fundamental differences in the underlying biological processes (Figure 5C; Figure 5—figure supplement 2A). The AML sample with inferred-prenatal origin was enriched for many processes associated with development. In contrast, inferred-postnatal samples appeared more metabolism focused (Figure 5C; Figure 5—figure supplement 2A). This further supports the association of these eRegulons with presumed prenatal origin t(8;21) AML, compared to postnatal origin disease.

Previous analysis using the TARGET (Bolouri et al., 2018) and BeatAML (Burd et al., 2020) datasets indicated that inferred-prenatal and inferred-postnatal origin t(8;21) AML may harbor different levels of chemosensitivity based on published drug response signatures (Figure 4D). Therefore, we performed in silico perturbations of eRegulon-associated TFs. PCA of the diagnosis and relapse samples recapitulated the expected differentiation trajectories along PC2, while separating diagnosis from relapse along PC1 (Figure 5D). Using the SCENIC+ (Bravo González-Blas et al., 2023) perturbation simulation workflow, we identified TFs estimated to induce differentiation, as defined by a negative shift in PC2 (Figure 5—figure supplement 2B). We prioritized TFs predicted to impact the HSC compartment and identified 18 TFs with predicted significant effects on HSC differentiation (Figure 5—figure supplement 2C). Several of these are components of the AP-1 complex (JUN, ATF4, FOSL2), which are established downstream targets of the t(8;21) fusion protein and are known to propagate t(8;21) AML (Ptasinska et al., 2019; Eferl and Wagner, 2003; Figure 5E; Figure 5—figure supplement 2C).

Using AP-1 complex members as a comparative baseline, we identified EP300 as one of the most impactful hits. EP300 has recently been shown to drive t(8;21) AML self-renewal through acetylation dependent mechanism (Wang et al., 2011). This suggests that presumed prenatal origin pediatric t(8;21) AML may be particularly sensitive to EP300 inhibition. One of the most striking predictions, for both diagnostic and relapse HSC populations, was BCLAF1 (Figure 5E; Figure 5—figure supplement 2C). BCLAF1 is a regulator of normal HSPCs (Crowley et al., 2022a), and its expression level declines during hematopoietic differentiation. While recent studies have identified a role for BCLAF1 in AML (Dell’Aversana et al., 2017), this has not been explored in detail in the context of pediatric AML or t(8;21) AML and may present a therapeutic opportunity.

We also performed SCENIC+ (Bravo González-Blas et al., 2023) perturbation modelling on the postnatal origin sample (AML12). In this case, the PCA was less straightforward to interpret, as branching differentiation trajectories towards a lymphoid or myeloid fate appear along the PC2 axis, while PC1 distinguishes diagnosis and relapse samples (Figure 5—figure supplement 2D). Therefore, we prioritized TFs based on a predicted effect similar to AP-1 complex components, as it is known that this complex is a critical regulator in t(8;21) AML. We identified several TFs from our original postnatal origin signature were predicted to have an effect (Figure 5—figure supplement 2D–F), supporting the relevance of the GRNs identified in our previous analyses.

To further investigate EP300 and BCLAF1, we queried the DepMap (Tsherniak et al., 2017) database to assess the dependency of t(8;21) AML cell lines to these genes (Figure 5—figure supplement 2G). We found that the two widely used cell lines of t(8;21) AML, KASUMI-1 (7-year-old donor) (Asou et al., 1991) and SKNO-1 (22-year-old donor) (Matozaki et al., 1995), were among the most sensitive to these perturbations based on their DepMap effect scores (Figure 5—figure supplement 2G). Several other cell lines sensitive to BCLAF1 were derived from pediatric cancers, most notably neuroblastomas, which also arise in-utero (Körber et al., 2023; Figure 5—figure supplement 2H). Together, these findings suggest that EP300 inhibition may be particularly effective in t(8;21) AML, and that BCLAF1 may present a new therapeutic target for t(8;21) AML, particularly in pediatric cases with inferred prenatal origin of the driver translocation.

Overall, our study demonstrates that large-scale single-cell data integration is a powerful approach to dissect specific patient groups in detail and enables robust comparative analyses. We present the AML scAtlas as a publicly available resource for the research community to address diverse biological questions. By applying AML scAtlas to t(8;21) AML, we identified age-associated gene regulatory networks that likely reflect differences in the developmental origins, biology, and outcome of the disease. These findings also highlight novel candidate therapeutic targets which may be more relevant in pediatric t(8;21) AML compared to adult-onset disease, offering opportunities for more tailored treatment strategies.

A literature search was performed for published AML scRNA-seq datasets (Velten et al., 2021; van Galen et al., 2019; Beneyto-Calabuig et al., 2023; Stetson et al., 2021; Zheng et al., 2017; Petti et al., 2019; Jiang et al., 2020; Johnston et al., 2020; Pei et al., 2020; Li et al., 2023; Lasry et al., 2023; Fiskus et al., 2023; Naldini et al., 2023; Mumme et al., 2023; Zhang et al., 2023; Li et al., 2022; Oetjen et al., 2018; Setty et al., 2019; Caron et al., 2020). Suitable studies were selected based on the data quality (over 1000 counts and 500 genes detected per cell for most of the data). Diagnostic, primary AML samples were selected from each AML study. Where healthy donor samples were present, these were also included, along with an additional 4 studies with healthy bone marrow samples.

Each scRNA-seq dataset underwent initial quality control individually using Scanpy (v1.9.3) (Wolf et al., 2018) as some studies provided raw data and others provided pre-filtered data. Where raw data was provided, doublets were removed using Scrublet (v0.2.3) (Wolock et al., 2019) and cells were filtered using the median absolute deviation as described in this single-cell best practices handbook (Heumos et al., 2023; Heumos and Schaar, 2023).

Once filtered, datasets were combined, and quality control was performed using Scanpy (v1.9.3) (Wolf et al., 2018). The full dataset had quality thresholds applied (percentage mitochondrial counts <10, read counts >1000, gene counts >500), removing any samples which had fewer than 50 cells remaining after filtering. Genes present in <50 cells were removed. MALAT1 was removed as this was highly abundant in many cells and considered artefactual.

The presence of batch effects was determined through dimensionality reduction and clustering using Scanpy (v1.9.3) (Wolf et al., 2018) and using the kBET algorithm (v0.99.6) (Büttner et al., 2019). This was repeated on individual studies, to assess whether there were sample-wise batch effects. Batch correction benchmarking was implemented using Harmony (Scanpy v1.9.3 implementation) (Korsunsky et al., 2019), scVI (v1.0.3) (Lopez et al., 2018), and scANVI (v1.0.3) (Xu et al., 2021) and quantified using scIB (1.1.4) (Luecken et al., 2022). Different numbers of highly variable genes were used to select the optimal number for integration. Batch correction was performed using scVI (v1.0.3) (Lopez et al., 2018) with the top 2000 highly variable genes, using sample as the model covariate.

The scVI corrected embedding was used to run UMAP and Leiden clustering using Scanpy functions (v1.9.3) (Wolf et al., 2018). Cell type annotation was performed using CellTypist (v1.6.0) (Domínguez Conde et al., 2022) using the ‘Immune_All_Low.pkl’ model, SingleR (v2.0.0) (Aran et al., 2019) using the Novershtern hematopoietic reference (Novershtern et al., 2011), and scType (v1.0) (Ianevski et al., 2022) with the tissue defined as ‘Immune system.’ Full automated tool outputs are detailed in Figure 1—source data 2; overall, we found that the results given varied significantly between different tools. We postulate that this is, in part, due to differences in the reference profiles used. Thus, we opted to use the best consensus of these different tools for our cluster identity assignments.

HSPC clusters were selected from AML scAtlas, and the scVI corrected embedding was used to recompute UMAP using Scanpy functions (v1.9.3) (Wolf et al., 2018). As our previous cell type annotations used generic reference profiles and were not AML-specific, we generated a custom cell type annotation reference to identify LSCs. We created a custom SingleR (v2.0.0) (Aran et al., 2019) reference using the Zeng et al., 2022 revised annotations of the van Galen et al., 2019 dataset (Figure 1—source data 2). This was also correlated with the LSC6 (Elsayed et al., 2020) and LSC17 (Ng et al., 2016) scores for each cell. To compare LSC abundance between ELN risk groups, chi2_contingency was implemented from SciPy (v1.12.0).

Samples with the t(8;21) translocation were selected from the full AML scAtlas. The UMAP was re-computed, and genes were filtered to remove those detected in fewer than 50 cells for the revised dataset, leaving 24,866 genes remaining. Gene regulatory network analysis was performed using pySCENIC (Aibar et al., 2017; Van de Sande et al., 2020) (v0.12.1) as per the recommended workflow. To facilitate comparisons between age groups, cell types were focused on HSPCs, as many adult samples were originally enriched for CD34. The RSS was calculated for the adult and pediatric samples to select the top 20 differential regulons per age group. Using SciPy hierarchical clustering (v1.12.0), regulons were filtered to identify regulon signature groups used for downstream analysis.

Bulk RNA-Seq data was downloaded for the TARGET (Bolouri et al., 2018) and BeatAML (Burd et al., 2020) cohorts and samples with t(8;21) were selected. Only bone marrow samples taken at diagnosis were used for downstream analyses (Source data 2). Using the previously defined signature regulons, the AUCell algorithm (Aibar et al., 2017) (v0.12.1) was implemented to measure regulon activity. Hierarchical clustering was performed using SciPy v1.12.0 to identify samples most enriched for each age-related signature. Differential gene expression analysis was implemented using edgeR (Robinson et al., 2010) (v3.42.4) and DESeq2 (Love et al., 2014) (v1.40.2), using a log2 fold change threshold of 0.5 and an adjusted p-value cutoff of 0.01. Candidate differential genes, ranked on log2 fold change, underwent GSEA with GSEApy (v0.10.8) using a significance threshold of 0.05. Cell type deconvolution was performed using AutoGeneS (Aliee and Theis, 2021) (v1.0.4) using the recommended workflow. Significance when comparing groups was ascertained using a Student’s t-test on the predicted cell type proportion values for each sample.

The Lambo et al., 2023 scRNA-seq and scATAC-seq data from pediatric AML bone marrow samples was downloaded, and the t(8;21) samples selected (Source data 2). Using our previously defined signature regulons, the AUCell algorithm (Aibar et al., 2017) (v0.12.1) was implemented to measure regulon activity. This identified the samples most enriched for each age-related signature as AML16 and AML12.

The SCENIC+ pipeline (Bravo González-Blas et al., 2023) (v1.0a1) was implemented as per the recommended Snakemake workflow for creating pseudo-multiome data. Regulons were filtered by correlation between modalities, using a threshold of 0.2 for non-multiome data. The most robust regulons were prioritized based on the SCENIC+ recommendations (direct +/+) (Bravo González-Blas et al., 2023). To facilitate comparisons, the eRegulon RSS was calculated for each patient and the top 30 eRegulons selected. The correlation between the gene sets underpinning eRegulons was calculated and sample-associated clusters were selected for over-representation analysis with clusterProfiler (Wu et al., 2021) (v4.8.3).

To predict the impact of specific TF perturbations on key cell types, SCENIC+ (Bravo González-Blas et al., 2023) perturbation modelling was implemented using the recommended parameters. TFs were then prioritized on their predicted impact on HSC differentiation and visualized using the PCA embedding. Candidate targets EP300 and BCLAF1 were queried in the DepMap (Tsherniak et al., 2017) databases to infer their potential importance.