+

Main

The basis for understanding and defining human pathophysiological states is a detailed description of interindividual heterogeneity among healthy individuals. Large population studies have identified wide interindividual differences in complete blood counts (CBCs) of healthy individuals1 and exposed different age-related blood count changes, such as high red blood cell (RBC) distribution width (RDW), macrocytic anemia and a reduction in absolute lymphocyte counts2. The establishment of reference values, or population-wide normal ranges for certain blood parameters, has been crucial for patient evaluation, diagnosis and treatment.

Although CBC reference ranges are used in the clinic daily, the equivalent reference range for hematopoietic stem and progenitor cells (HSPCs) has not been established so far. As HSPCs reside mainly in the bone marrow (BM), access to these cells, especially in the healthy population, has been problematic, whereas their general paucity in the circulation made it quite challenging to characterize them efficiently from the blood. This has become feasible given modern technologies such as single-cell RNA sequencing (scRNA-seq).

Individual heterogeneity in the frequency of circulating HSPCs (cHSPCs) has been reported in the past and was linked to age, sex, smoking status, lipid profiles and hereditary factors3, as well as to different pathological states4. Few studies have analyzed HSPC heterogeneity in higher resolution, but their sample size was limited5,6. Previous studies, including some based on scRNA-seq analysis7, demonstrated that most HSPC subpopulations can be identified in the peripheral blood (PB)8 and functional stem cells were identified in the PB of mice9 and humans7.

We have developed a reference model for healthy cHSPC distributions and provided proof-of-concept evidence supporting potential diagnostic applications. We applied scRNA-seq analysis to cHSPCs from 148 healthy age- and sex-diverse individuals, to capture a spectrum of states, from hematopoietic stem cells (HSCs), through early common myeloid and lymphoid progenitor states and more specific progenitor populations. All data can be explored in https://apps.tanaylab.com/MCV/blood_aging. We discovered extensive interindividual heterogeneity in the frequency of cHSPC subtypes and found that these correlate with certain CBC parameters, aging and the presence of clonal hematopoiesis (CH). We then developed tools for projecting new samples on our reference model and analyzed 73 additional samples from patients with cytopenia and myelodysplastic syndrome (MDS), to demonstrate our model’s potential applications in MDS diagnosis. The healthy reference model and methodologies used in the present study provide a framework for the deployment of single-cell genomics in hematology, for the diagnosis of MDS, and possibly other stem cell-related blood malignancies, from the PB, reducing the need for BM analysis.

Results

HSPC states observed across humans in PB

To evaluate interpersonal diversity in subtype distribution and regulation of cHSPCs, we combined multiplexed scRNA-seq, bulk DNA genotyping and integrated clinical data (Fig. 1a). Multiplexing was resolved using SNPs identified in the 3′-UTR of cHSPCs’ RNA, facilitating precise matching of cells to individuals and improving control for batch effects and doublets. Altogether, we collected cHSPCs from 79 men and 69 women between the ages of 23 years and 91 years (median 61.5 years) (Extended Data Fig. 1a and Supplementary Table 1). We performed deep targeted somatic mutation analysis to identify cases of CH (Supplementary Table 1)10. After quality control and filtering, we retained 840,104 single-cell profiles, which were normalized to control for sequencing-platform batch effects and combined to construct and annotate a metacell manifold model11 (Extended Data Fig. 1b,c). We retained 626,966 CD34+ single cells for downstream analysis (Extended Data Fig. 1d). These formed a rich repertoire of states, associated with cHSPCs and their differentiation trajectories (Fig. 1b and Extended Data Fig. 1e,f). The derived model recapitulated and deepened earlier characterization efforts of HSPC states from the BM. We noted that, although we could not assume that cHSPCs fully reflect BM HSPC dynamics, previous studies, as well as our own BM scRNA-seq comparisons, supported at least partial compatibility between the two12 (Extended Data Figs. 2 and 3a). One notable characteristic specific to cHSPCs was, however, the repression of cell-cycle gene expression (Extended Data Fig. 3b), previously demonstrated by others7. Importantly, we found our cHSPC model to be consistent across individuals. The median number of individuals contributing cells to each metacell was 84 and all metacells included cells from at least 47 individuals. Expression differences between cell states were greater than between individuals, limiting individually specific differential expression when controlling for each sample’s cell distribution over the atlas states (Extended Data Fig. 3c,d). Altogether, the data suggest that cHSPCs, although not fully reflecting BM hematopoiesis, can serve as a highly accessible proxy for hematopoietic dynamics.

Fig. 1: Mapping cHSPCs.
Fig. 1: Mapping cHSPCs.
Full size image

a, Experimental design. b, Annotated two-dimensional Uniform Manifold Approximation and Projection (UMAP) of our metacell manifold after filtration of metacells with low CD34 expression. For all subsequent panels in Figs. 1–3, metacell color denotes cell state as here. c,d, Symmetrical (c) and asymmetrical (d) regulation of specific HSC TFs on bifurcation to the CLP (right) and MEBEMP (left) lineages. Each panel shows the expression of one gene (y axis). Metacells in all panels are ordered (left to right) by increasing AVP expression in the MEBEMP lineage and decreasing AVP expression in the CLP lineage. The y axes denote log2(fractional expression) of each gene. e, The metacell population of interest (dashed line) linking BEMPs to their MEBEMP-L precursors. f, Positively and negatively regulated TFs involved in early BEMP differentiation. g, Gene–gene plot of IRF8 against TCF7 expression as Hallmark markers of DC and T cell differentiation, respectively. The high ACY3 NKTDP metacell population of interest is depicted (dashed line). h, This population exhibits high expression of both T cell and DC regulators, forming a gradient consisting of NK or T cell-like progenitors exhibiting a high TCF7:IRF8 expression ratio along with high expression of other T cell Hallmarks such as CD7, MAF, IL7R, TRBC2 and DC-like progenitors exhibiting a low TCF7:IRF8 expression ratio, along with high expression of other DC Hallmarks, such as the myeloid TF PU.1 and the MHC-II gene CD74. Panel a created with BioRender.com.

+ HLF, GATA3, HOXB5 and TLE4 as HSC TFs

One of the hallmarks of our cHSPC model is a distinct HSC state that is transcriptionally linked with two major differentiation gradients: the first represents a continuum of common lymphoid progenitors (CLPs; subdivided into early (E) mid (M) and late (L) states). The second, more common branch, represents multipotent progenitor (MPP) states and their differentiation toward granulocyte–monocyte progenitors (GMPs), erythrocyte progenitors (ERYPs) and basophil, eosinophil or mast progenitors (BEMPs). Platelet contamination prevented precise megakaryocyte progenitor modeling (Extended Data Fig. 3e), such that states at the base of the myeloid trajectory were annotated as megakaryocyte, erythrocyte, basophil, eosinophil or mast progenitors (MEBEMPs, subdivided into early (E) and late (L) states).

Early HSCs are marked by high AVP and HLF expression and were previously shown to represent a rare cell population enriched with self-renewal capacity in both the BM and cord blood13. Our model included data on 14,440 HLF and AVP expressing HSCs that could be matched with cells from independent BM atlases14, suggesting that, under a steady state, HSCs with potential self-renewal capacity are present in the PB (Extended Data Fig. 3f). Further functional studies are needed to establish this finding. Together with HLF and AVP, we discovered 14 genes expressed at least 1.75-fold higher in HSCs compared with their 2 immediate differentiation branches (Extended Data Fig. 3g and Supplementary Table 2). We identified several transcription factors (TFs) enriched in HSCs, including HOXB5, TLE4 and GATA3 (Fig. 1c). GATA3 was previously reported to regulate self-renewal in murine long-term HSCs15. Its role in human HSCs has not been studied thus far. We note that, although the HSC state is defined by unique markers that are symmetrically downregulated on exiting to the CLP and MEBEMP trajectories (Fig. 1c), it also expresses several lineage-specific regulators at intermediate levels, which are bifurcating anti-symmetrically on exiting the HSC state to the CLP and MEBEMP trajectories (Fig. 1d and Extended Data Fig. 3g). This may suggest that the multipotent capacity of HSCs is associated with intermediate expression of multiple regulators, which is resolved with differentiation.

BEMPs and NKTDPs are enriched in cHSPCs

The cHSPC atlas was enriched for BEMPs. Although classic studies linked these cells with a GMP origin, more recent studies suggested that these emerge, at least in part, from erythroid progenitors in both mice and humans7,16. Our analysis identified a small population of metacells linking BEMPs with their MEBEMP-L precursors (Fig. 1e). This highlighted TFs (Fig. 1f) and other factors (Extended Data Fig. 4a) positively or negatively regulated in this postulated early stage of BEMP specification. Another interesting cHSPC population included lymphoid states with high ACY3 expression and intermediate-to-low DNTT levels, a combination rarely found in human BM but present in PB (Extended Data Fig. 4b). We observed co-variation of key T cell regulators within this population and anti-correlation of these factors with some Hallmark plasmacytoid dendritic cell (pDC) regulators, as demonstrated by comparison of TCF7 and IRF8 expression (Fig. 1g,h and Extended Data Fig. 4c). We therefore termed this population natural killer (NK) cell, T cell and DC progenitors (NKTDPs)17,18. To summarize, our map of cHSPCs showed a rich spectrum of progenitor states, which refined previous analyses and a remarkable consistency of these states across individuals. This provided an opportunity for deciphering interindividual hematopoietic variability based on our solid and quantitative definition of cHSPC subtypes.

Interindividual variation in cHSPC state composition

To study interindividual cHSPC variation, we first analyzed cell-state compositions by quantifying cell-state relative frequencies within each individual’s single-cell ensemble (Fig. 2a). These frequencies varied extensively between individuals as shown in Fig. 2b. For example, HSCs and CLP-Ms, representing 2.4% and 12.6% of the CD34+ population on average, showed s.d. values of 1.0% and 6.8%, respectively. The abundant MPP and MEBEMP-E states (mean frequencies of 20.7% and 37.6%) showed smaller relative variations (s.d. values of 4.9% and 5.8%, respectively). To analyze the stability of cell-state frequencies across time and sampling instances, we re-sampled 20 individuals 1 year after their original sampling date. Both CLP (CLP-E, CLP-M, CLP-L and NKTDP) and MEBEMP (MEBEMP-E, MEBEMP-L, ERYP and BEMP) frequencies were stable within the same individual across time (Fig. 2c).

Fig. 2: Normal cHSPC composition.
Fig. 2: Normal cHSPC composition.
Full size image

a, Characterization of interindividual cHSPC compositional variation and its correlation to clinical parameters (scheme). b, Boxplots of cHSPC state frequency distributions across 148 healthy individuals (logarithmic scale). The percentage was calculated from all CD34+ cells within each individual’s single-cell ensemble. Boxplot centers, hinges and whiskers represent the median, first and third quartiles and 1.5× the interquartile range, respectively. Outliers are marked by circles. The numbers represent the mean ± s.d. for each distribution. c, Comparison of cell-state frequencies between 19 biological replicates and their original samples, for CLP (CLP-E, CLP-M, CLP-L and NKTDP) populations (top) and MEBEMP (MEBEMP-E, MEBEMP-L, ERYP and BEMP) populations (bottom). The diagonal y = x is shown in red. All biological replicates were sampled 1 year after their original sampling date. d, Top: cell-state frequency profiles over the HSC-MEBEMP and HSC-CLP differentiation gradients of six sampled individuals (colored lines), each representing one of six archetypes (classes) of cHSPC composition observed in healthy individuals. The dashed lines represent the median (black) and the 5th and 95th percentiles (gray) of the studied population. Bottom: cell-state enrichment map over 15 differentiation bins (rows), for all studied individuals (columns) clustered into six classes (Methods). Classes I and II represent individuals relatively enriched in lymphoid progenitors, whereas classes V and VI represent individuals with relative depletion of lymphoid progenitors. Individuals are sorted by stemness within each class. Age and sex are denoted for each individual. e, CBC correlations to cell-state frequencies: %lymphocytes (from white blood cells, calculated for the entire cohort, left), HCT (men only, center) and RDW (men only, right). Missing individuals lacked sufficient cells for analysis. Two-sided permutation test P values are displayed for each correlation. See Methods for details on the permutation-based test. f, CH frequency (by gene) in age- and sex-matched high (red, n = 602) and normal (black, n = 602) RDW individuals selected from a cohort of 18,147 individuals. Panel a created with BioRender.com.

To analyze composition in higher resolution, we profiled each individual’s enrichment over the CLP and MEBEMP trajectories. Clustering of these enrichment profiles yielded six archetypes of cHSPC composition within the healthy population (classes I–VI, Fig. 2d). These were composed of individuals with relative lymphoid enrichment (classes I and II) or depletion (classes V and VI), further subdivided by a stemness gradient, enriched in classes II, IV and VI and depleted in classes I, III and V. Analysis of technical and biological replicates confirmed this variation to be robust and individual-specific (Extended Data Fig. 5a,b). To summarize, we constructed cHSPC subtype normal reference ranges and showed that, although HSPC cell states are consistent among healthy individuals, their compositions are highly variable.

Circulating HSPC frequencies correlate with CBCs and CH

To extract an initial clinical annotation for the observed interindividual variation in cHSPC state frequencies, we correlated individual compositions with longitudinal CBC data (Methods). We observed a significant positive correlation (P < 0.01) between PB mature lymphocyte counts (%) and CLP frequencies (Fig. 2e, left). Given the high variability in female RBC counts and volumes during menstruation, pregnancies and prolonged perimenopausal periods, we analyzed RBC indices (count, hematocrit (HCT), mean corpuscular volume (MCV) and RDW) separately for men and women. We observed a significant negative correlation (P < 0.02) between CLP frequencies and HCT (men, Fig. 2e, middle) and a significant positive correlation (P < 0.01) between increased RDW and relative CLP depletion (men, Fig. 2e, right). Our previous work19 and the work of others20 correlated high RDW with CH and predisposition to acute myeloid leukemia. Our data suggest that reduction in CLP frequencies is associated with CH (Extended Data Fig. 6a). A similar trend was suggested by genotyping of transcriptomes (GoT)21 performed on one of our DNMT3A R882H cases, showing a lower fraction of CLP cells within the mutant clone (P < 0.005, Fisher’s exact test; Extended Data Fig. 6b). Although this trend was suggested in other GoT data22, sample size is insufficient to prove it statistically and explore the clonal mechanisms underlying it. To further explore the association between CH and RDW, we studied a cohort of 18,147 healthy individuals for whom we had both longitudinal CBCs and DNA available. We identified 602 individuals with a high RDW (>15%, not meeting minimal criteria for MDS diagnosis) and 602 age- and sex-matched normal RDW controls. We performed deep targeted sequencing to identify leukemia-associated mutations on both high-RDW individuals and controls, and found a significant enrichment of CH+ cases in the high-RDW group (Fisher’s exact test, P < 0.002; Fig. 2f and Supplementary Tables 3 and 4). Altogether, the data suggest associations across decreased CLP frequencies and elevated RDW and CH. Determination of the existence of a direct (and perhaps three-way) linkage for these variables requires further investigation.

Age-related myeloid bias is predominantly observed in men

Analysis of age-linked compositional changes in cHSPCs within CH-negative individuals showed a remarkable increase in myeloid (MEBEMP) to lymphoid (CLP) progenitor ratios in men (when comparing <50 to >60-year-old individuals; Fig. 3a and Extended Data Fig. 6c). This effect was insignificant in women. Of note, although both men and women experience a decline in lymphocyte counts with aging, it occurs at an older age in women, as confirmed by recent analyses2. Notably, women show a temporary postmenopausal surge in lymphocyte counts, which delays their decline. Within the MEBEMP differentiation trajectory, aging was correlated with over-representation of more differentiated states, once again only in men (Fig. 3b). Of note, the frequency of cHSCs did not significantly change with age (Fig. 3c). Although previous studies suggest that aging is linked with an increase in HSC frequency23,24, this was not observed with the restrictive definitions employed in the present study. We further identified an age-related decline in CD34+ HSPC frequency in a cohort of 1,000 healthy individuals undergoing peripheral blood mononuclear cell (PBMC) scRNA-seq25 (Fig. 3d and Extended Data Fig. 6d), which has also been reported by fluorescence-activated cell sorting (FACS) in the past3. The sex-specific correlation between age and cHSPC myeloid bias could be related to cell intrinsic properties, such as male-specific leukemia-associated mutations predisposing to myeloid differentiation26. This is less likely because canonical CH-positive cases were excluded from this analysis. Alternatively, this predominantly male myeloid bias could be related to cell extrinsic factors such as age-related hormonal and BM microenvironmental changes27,28.

Fig. 3: Age- and sex-linked changes in cHSPC composition.
Fig. 3: Age- and sex-linked changes in cHSPC composition.
Full size image

a, Frequency of MEBEMP (MEBEMP-E, MEBEMP-L, ERYP and BEMP, left) and CLP (CLP-E, CLP-M and CLP-L, right) populations, out of a total CD34+ population, in young (<50 years) versus old (>60 years) individuals without CH, in men (blue) and women (red). The two-sided Kruskal–Wallis P values for differences among groups are denoted. The number of individuals per group is (left to right): 31, 15, 24 and 31. b, Analysis of age-linked compositional differences within the MEBEMP differentiation trajectory, comparing abundance of more (MEBEMP-L) with less (MPP) differentiated states in young versus old individuals. The two-sided Kruskal–Wallis P value for difference among groups is denoted. The number of individuals is as in a. c, As in a, for the HSC population. d, cHSPC frequency per age decimal in an scRNA-seq PBMC dataset of 1,000 healthy individuals25. For each decade, mean CD34+ cell frequency is shown (Methods). The 95% confidence intervals are indicated as error bars. The number of individuals in each decade is indicated (top). e, True age (x axis) versus age predicted based on composition-controlled MPP expression (y axis). The diagonal y = x is shown in red. f, Gene–gene correlation heatmap, calculated over individual-level MPP gene expression controlled for MPP composition. g, Individual LMNA signatures (log2(observed:expected ratios)) in lymphoid (CLPs) and early myeloid (MPPs) cell states. h, Analysis of age-linked differences in LMNA signature expression for CLP (right) and MPP (left) populations in young (<50 years) versus old (>60 years) individuals. The y axis denotes log2(observed:expected expression) normalized for composition. The number of individuals is 66 young, 65 old on the left and 48 young, 53 old on the right. The two-sided Mann–Whitney U-test P values are indicated. i, Individual heatmaps of single-cell counts over 20 bins of stemness (AVP signature, y axis) and MEBEMP differentiation (GATA1 signature, x axis). Individual identifier, MCV and RBC are denoted at the top. The diagonal is indicated in black for reference. j, MCV versus RBC in male donors, with colors indicating high (red) and low (black) sync-scores. k, Correlation between individual sync-scores and cell-state compositions in men. The two-sided permutation test P value is denoted. All boxplots are as in Fig. 2b.

Composition-controlled cHSPC expression correlates with age

As shown above, individual cHSPC compositions provide an initial blueprint of hematopoietic dynamics along the stemness and CLP or MEBEMP axes, with age-dependent changes. Composition-normalized gene expression profiles were further correlated with age, enabling age prediction based on normalized gene expression alone (Fig. 3e and Extended Data Fig. 6e; see Supplementary Tables 5 and 6 for additional screening for age-, CBC-, CH- and sex-associated gene expression). We next looked for gene groups (signatures) that co-variate between individuals, filtering out sex-linked signatures and those showing strong batch effects. The most prominent of these signatures included Lamin-A (LMNA) as well as ANXA1, AHNAK, MYADM, TSPAN2 and VIM, among others (Fig. 3f, Extended Data Fig. 6f and Supplementary Table 7). Individual LMNA signature expression varied across a range of more than twofold (Extended Data Fig. 6g), exhibiting high expression variability in HSCs and early myeloid and lymphoid cell states, and a homogeneously low expression in late MEBEMPs and CLPs (Extended Data Fig. 6h). Individual LMNA signature expression was consistent in myeloid and lymphoid cell states (Fig. 3g) and was stable in our follow-up cohort (Extended Data Fig. 6i). We observed an age-linked increase in LMNA signature expression in lymphoid, but not myeloid, cHSPCs (Fig. 3h). Future studies on larger cohorts, enriched with clinical data, could further explore the age-related LMNA signature overexpression in CLPs and how it correlates with disease and immune function. Taken together, we show that, in addition to the accumulation of leukemic mutations in HSPCs, aging is linked with changes in the distribution of progenitor cell states within the PB and with notable expression differences in certain gene signatures. The mechanistic basis for this variation and its clinical impact remain unresolved.

Coordination of stemness and myeloid signatures

The differentiation of HSPCs toward MEBEMP and CLP fates involves coordinated activation and repression of specific transcriptional programs which are conserved across individuals. Yet, our screen for interindividual variation in gene signature expression suggested that individuals differed in the way in which they synchronized the opposing effects of these stemness and differentiation programs. To quantify this variation, we compared AVP (stemness) and GATA1 (MEBEMP differentiation) signatures (Supplementary Table 8) on a 20 × 20 bin expression matrix (Fig. 3i). Although most individuals displayed dynamics close to the diagonal line (individuals N16 and N86, for example), following the typical transition from stemness to differentiation, some individuals deviated from the diagonal, indicating skewed synchronization between AVP and GATA1 signatures. We quantified this deviation (that is, off-diagonal frequency) using a synchronization-score (sync-score). This facilitated the identification of individuals with sync-scores as low as 0.12 (N122 and N172, for example, Fig. 3i, top), indicating delayed activation of GATA1 relative to AVP repression. In contrast, individuals exhibiting a high sync-score (N98 and N121, for example, Fig. 3i, bottom) show early activation of GATA1 expression, which precedes AVP repression. Interindividual sync-score variability (Extended Data Fig. 6j) was positively correlated with RBC levels and consistently anti-correlated with MCV in men (P < 0.01 (Spearman’s) for both RBC and MCV; Fig. 3j). Analysis of the correlation between individual sync-scores and cHSPC compositions in men demonstrated a negative correlation with ERYPs (Fig. 3k). In summary, variation in the coordination of stemness and MEBEMP differentiation programs correlated with RBC counts and volumes. More studies on larger cohorts are needed to explore how this coordination relates to age-related macrocytic anemia.

Circulating HSPC composition abnormality in cytopenia and MDS

Diagnosis of myeloid malignancies requires the identification of clonal markers (mutations or structural variants) and the detection and quantification of blasts and dysplasia, by next-generation sequencing, polymerase chain reaction, cytogenetics, fluorescence in situ hybridization, microscopy and flow cytometry of BM specimens. In Fig. 4a we described a stepwise approach for analysis of myeloid disorders based on sampling of cHSPCs and comparison of their compositions, normalized expression and copy number variations (CNVs) to our normal reference (Extended Data Fig. 7a–c). As proof of concept, we focused on MDS diagnosis. First, we reconstructed the reference model using data from 79 healthy individuals, putting aside some normal samples for classifier training. We then performed additional sequencing to obtain data from 44 patients with MDS and 29 patients with cytopenia (Supplementary Tables 9–11). We developed a streamlined in silico sorting scheme for quantifying the cHSPC composition of a new PB sample given the reference (Extended Data Fig. 7a,b) and used it to identify cases with abnormal compositions (Fig. 4b,c, Extended Data Fig. 8a and Methods). Classification included subpopulations (GMP-L, pre-B, pro-B and MkP) that were rare in the normal reference model and were not shown in Fig. 1. We then marked MDS or cytopenic samples with normal compositions (matching the reference model, group 1) and organized them along the myeloid and lymphoid spectrum. The remaining cases were clustered into distinctive subclasses. Although most cases of MDS showed significantly lower CLP frequencies (groups 3 and 4; Extended Data Fig. 8b), we identified a subclass of MDS and cytopenia with high CLP frequencies (group 2). Other subclasses included high MPP (group 4.2), high BEMP (group 4.1) and high GMP (group 3) frequencies. This sorting scheme partially separated MDS from other, non-MDS-related, cases of cytopenia, with most cytopenia cases exhibiting normal (group 1) compositions. Cases of MDS with abnormal CNVs (Methods) were enriched in groups 2–4 (P < 0.004, Fisher’s exact test; Fig. 4d and Supplementary Table 12) and patients with high RDW were enriched in group 4 (Extended Data Fig. 8c). In summary, cHSPC compositions reveal molecular features that offer possibilities for identifying MDS subclasses and pathophysiology. Classification of MDS cases with normal cHSPC compositions (group 1) depends on further analysis of genetic and transcriptional states within specific cHSPC subtypes.

Fig. 4: Applying the cHSPC reference model to MDS diagnosis and subclassification.
Fig. 4: Applying the cHSPC reference model to MDS diagnosis and subclassification.
Full size image

a, Schematics of a diagnostic approach to cytopenia and MDS using scRNA-seq of cHSPCs and a reference model. b–d, Cytopenia and MDS patient cHSPC compositions and mutations. b, Each bar represents a patient’s cHSPC composition. Patients exhibiting normal compositions (Methods) are ordered by lymphoid cHSPC frequency (left) and those exhibiting abnormal compositions are ordered by composition hierarchical clustering (right). Cytopenia or MDS subclasses suggested by composition are marked (top). c, Patient composition abnormality scores, depicted by both bar height and color, as coded on the right. d, Patient diagnosis and CH VAF for three specific mutations and the maximal VAF over all other detected mutations, as color coded on the right. Presence of copy number alterations (CNAs) as detected by scRNA, color coded in black. e, Transcriptional signatures across healthy donors and patient groups as in b, further subdivided by sex (men, left; women, right). Individuals exhibiting insufficient cell counts for the population of interest were excluded. Color denotes clinical diagnosis. Mann–Whitney Benjamini–Hochberg-adjusted significance of difference from healthy donors of the same sex is indicated by asterisks (*q < 0.05, **q < 0.01, ***q < 0.001). f, Receiver operator curve for a classification model predicting MDS status based on cHSPC scRNA-seq, CH VAF and CBC values. FPR, false-positive rate; TPR, true-positive rate. g, Frequency of a CLP-E-like cell state across healthy donors and cytopenia and MDS groups, as in e. h, Comparison of BM blast counts measured by FACS and frequency of a CLP-E-like cHSPC population across individuals. Color denotes clinical diagnosis, as in e. Samples excluded from b–g due to the presence of a complex karyotype (as detected by scRNA-seq) are highlighted (arrows). Linear fit across all individuals (n = 26) is shown (dashed line), as well as the corresponding r and (two-sided) P value.

PB-based MDS diagnosis with CBC, mutation and cHSPC RNA data

To improve our diagnostic accuracy, we next derived specific gene signatures showing additional variation within cell types, from the reference model (Supplementary Table 13), and scored these signatures based on their ability to separate patients with MDS and cytopenia from healthy donors. A group of major histocompatibility complex class II (MHC-II) genes in MEBEMP-L, multi-potency genes in BEMP and S-phase genes in MEBEMP-L (Fig. 4e) emerged as top-ranking. These signatures were overall consistent across different samples of the same individuals (Extended Data Fig. 8d). We then combined CBCs, maximum variant allele frequency (VAF), cHSPC compositions and all afore-mentioned expression signatures into a feature set that formed the basis for construction of an MDS diagnostic classifier using standard machine learning tools. We created two cohorts: the first (cohort 1), composed of 28 patients with MDS, 20 patients with cytopenia and 41 healthy individuals, and the second (cohort 2) composed of 16 patients with MDS and 9 patients with cytopenia. We observed classifier training performance (even when aiming to separate MDS from cytopenia cases) was better when including normal cases in the dataset. Analysis of classifier performance showed very high specificity and sensitivity (Fig. 4f; area under the curve (AUC) = 0.93 in leave-one-out cross-validation for cohort 1 separation of MDS from cytopenia). Performance of the cohort 1 model on cohort 2 (which was not used during classifier training) was even higher (AUC = 0.97). Although cohort 2 data was not used in classifier training or feature selection, it was accessible to us during project analysis phases, such that we were cautious not to treat this as formal validation. The most informative feature used by the MDS classifier was the maximum VAF (Extended Data Fig. 9a). Yet, classifier performance was high even when excluding VAF information (Extended Data Fig. 9b,c).

Diagnosis and risk stratification of MDS rely on quantification of BM blast fractions. Our analysis of cohort 1 and cohort 2 samples, with the addition of three cases of MDS exhibiting complex karyotypes, suggests that we can predict this percentage quantitatively from cHSPC data using the fraction of cells showing a mixed HSC and CLP state (Fig. 4g,h and Extended Data Fig. 9d). All in all, this implies that, with further validation and testing, cHSPC profiling has the potential to replace BM analysis for MDS diagnosis and risk stratification, offering substantial benefits, such as noninvasive follow-ups and watchful waiting protocols. We present two case studies supporting this idea in Extended Data Fig. 10a,b. The first is an 82-year-old man showing progressive clonal expansion over a span of 3 years, accompanied by deteriorating anemia. The second is a 65-year-old woman presenting with clonal del5q showing complete cytogenetic remission after lenalidomide treatment. Additional follow-up examples (Extended Data Fig. 10c) suggest small changes in (normal or abnormal) composition across time, further supporting the idea of using cHSPCs for noninvasive assessment of disease progression.

Discussion

The present study characterizes interindividual heterogeneity in cHSPCs across 148 healthy individuals using scRNA-seq analysis of PB CD34+ cells. The magnitude of our cohort, along with the potency and resolution of modern single-cell technologies and the computational methods used in the present study, allowed us to characterize in detail the transcriptional programs of diverse, sometimes rare (NKTDP and BEMP), HSPC subpopulations, refining and augmenting previous findings from smaller cohorts (Fig. 1). We defined a normal reference range for cHSPC subpopulation frequencies within an age- and sex-diverse healthy population and showed that cHSPC subtype compositions were highly variable between individuals, whereas the cell states themselves were remarkably general (Fig. 2). These compositions remained stable over a 1-year follow-up period. Future studies will need to further explore and better define the mechanistic and genetic basis for this compositional heterogeneity. With current sample size, we showed that the known age-related myeloid bias in HSPCs is predominantly male driven and that composition-controlled RNA expression can be used to infer chronological age (Fig. 3).

Our data show that cHSPCs are transcriptionally similar to their BM counterparts (Extended Data Figs. 2 and 3), except for reduced cell-cycle gene expression. Although not a complete model for BM hematopoiesis, cHSPCs serve as a highly accessible proxy for key hematological processes. Interindividual differences in cHSPC compositions and states can thus serve as a tool for capturing key aspects of a patient’s hematopoietic state. The relevance and importance of a cHSPC normal reference (Fig. 2b) can perhaps be better understood in view of the normal CBC reference range, developed in the 1930s1. The development of a population-wide CBC reference enabled the identification of numerous pathological blood states that characterize distinct clinical entities. In a similar fashion, our cHSPC reference can be used to characterize physiological and pathological states. In Fig. 4 we describe a pipeline for the identification and characterization of blood pathologies based on our normal cHSPC reference and show how this can be applied to MDS diagnostics (including inference of cytogenetics and blast counts from the PB). We present scRNA-seq data on cHSPCs from 73 cases of cytopenia and MDS, greatly extending currently available BM MDS scRNA-seq datasets5,29,30,31,32. The data described supports MDS diagnosis (over non-MDS-related cytopenia) and suggests the possibility of MDS subclassification based on over-representation of distinct HSPC progenitor populations. The MDS-related gene expression signatures identified in the present study open avenues for research that might contribute to better understanding of MDS pathophysiology and drug design strategies. Importantly, further follow-up, validation in prospective studies and cohort expansion to ethnically diverse populations are needed to prove that the tools introduced here can become a clinical standard. The diagnostic potential of our reference model may be further enhanced upon acquisition and analysis of additional blood subpopulations and disease states in contrast with the reference. Practically, application of scRNA-seq for diagnostics would have to rely on stable and minimally biased cell acquisition and processing technologies that can be deployed across diverse clinical settings and provide consistent and trustworthy results. Development in this domain is promising, but more work must be done to reach clinical standards.

To conclude, our study delves into the basic molecular physiology of cHSPCs at the population level, uncovering age-related phenotypes and proposing a platform for mechanistic and diagnostic insights into blood malignancies. This resource, along with various other tools for profiling genetics and epigenomics in the blood, has the potential to redefine normal versus pathological states in hematology and provide both clinicians and researchers the means for mapping the transition from health to disease.

Methods

Patient recruitment

All healthy reference model individuals (n = 148, analyzed in Figs. 1–3 and Extended Data Figs. 1–6) volunteered to participate in our study and donated blood at the Weizmann Institute of Science (WIS) between November 2020 and December 2023. They were recruited from the WIS community and primary care clinics and consisted of 79 men and 69 women aged 23–91 (median 61.5) years. Their demographic data and CBCs are included in Supplementary Table 1. Written informed consent allowing access to their demographic, longitudinal CBC and sequencing data (CH and genotyping panels) was obtained from all participants in accordance with the Declaration of Helsinki. All relevant ethical regulations were followed and all protocols were approved by the WIS ethics committee (under Institutional Review Board (IRB) protocol no. 283-1).

For the main reference model (Figs. 1–3), recruitment was intended to allow characterization of the normal variation in cHSPC states. As no such profiling had been previously performed, we could not assume much about the variance in the population a priori. Participants were therefore required to lack any known hematological condition, including hematological malignancy or premalignant state, or any prior evidence of blood clonality. An Illumina-sequenced subset of these 148 individuals (n = 79) was used for constructing the healthy reference model used in Fig. 4 (‘Fig. 4 reference model’), filtering out individuals with any blood count abnormality (up to 5 years before sampling) and putting aside 41 healthy samples for classifier training.

Recruitment of the cytopenic cohort (including patients with MDS and non-MDS-related cytopenia, analyzed in Fig. 4 and Extended Data Figs. 7–10) took place between November 2021 and February 2024. These patients were recruited from several outpatient hematological clinics by collaborating physicians to represent the wide clinical spectrum of MDS, from patients with moderate anemia and mild dysplasia as their sole BM abnormality to those with severe cytopenia and excess blasts on BM analysis. Key patient characteristics, including CBCs, BM FACS blasts and mutational data, are included in Supplementary Tables 9 and 11. Median age for the cytopenic cohort was 73 years (range 27–93 years), with men representing 53% of patients. Patient PB samples were either drawn at WIS or transported to WIS within <2 h of blood drawing. All cytopenic samples were processed in an identical fashion to the healthy ones (described below). Longitudinal CBCs, mutational data and most recent BM analyses were collected from patients and analyzed along with their scRNA-seq data.

The cytopenic cohort included a total of 83 individuals, 50 of whom were labeled as cases of ‘MDS’, based on BM morphology and/or mutational and karyotypic abnormalities (as detected in the clinic or by our CH panel and scRNA CNA analysis). The remaining 33 patients with cytopenia not satisfying MDS criteria were labeled as cases of ‘cytopenia’. We note that, consistent with common medical practice in Israel, most of these 33 patients with cytopenia did not undergo BM examination, which may have resulted in missed MDS diagnoses. To address this limitation, we collected the most recent clinical data available for patients with cytopenia, with a median follow-up period exceeding 600 d after cHSPC sampling (Extended Data Fig. 10d and Supplementary Table 9). Importantly, no new diagnoses of myeloid malignancy were recorded in any of the cytopenic cases, except for N193 who was diagnosed with VEXAS syndrome 1 year after cHSPC sampling, exhibiting a UBA1 mutation c.121A>G;p.Met41Val, but had not been diagnosed with MDS yet. In addition, during this follow-up, median change in RDW% was zero. In contrast, over a similar period (looking at historical records before cHSPC sampling; Extended Data Fig. 10e), median RDW% change in cases of MDS was 0.75, significantly higher than in patients with cytopenia (Extended Data Fig. 10f; P = 0.001, two-sided Mann–Whitney U-test). Overall, these data support the accurate classification of cytopenic cases. Eleven cHSPC samples were acquired from patients under treatment, six of which were included in Fig. 4 (Supplementary Table 9). Of the 83 patients with cytopenia, 17 who presented with asymptomatic, mild cytopenia, were also included in the original healthy cohort of 148 individuals.

Sampling of cHSPCs

We drew 50 ml of PB from each individual into lithium–heparin tubes and put aside 1 ml of blood for DNA production. The remaining volume was used for PBMC isolation via Ficoll using Lymphoprep-filled Sepmate tubes (STEMCELL Technologies), followed by CD34 magnetic bead-based enrichment using the EasySep human CD34+ selection kit II (STEMCELL Technologies). We found this enrichment strategy to be simple and reproducible and chose it for a couple of reasons: (1) RNA-seq data were most reproducible when cells were not sorted, but rather enriched for using beads (lower mitochondrial gene fraction); and (2) CD34 purity could be highly regulated by this method, to achieve anywhere between 50% and 95% enrichment of CD34+ cells, which could later be easily distinguished based on their single-cell expression data.

ScRNA-seq of cHSPCs

ScRNA libraries were generated using the 10x Genomics scRNA-seq platform (Chromium Next Gem single-cell 3ʹ reagent kit v.3.1). Chip loading was preceded by flow cytometry to verify that enrichment was successful and that enough CD34+CD45int live cells were gathered. All blood samples were either drawn at WIS or transferred from participating clinics on the morning of each experiment day, and time from blood draw to 10x loading was restricted to 5 h. The motivation for working with fresh samples was based on our previous experience with PB CD34+ cells being vulnerable to freezing–thawing rounds and long manipulation times. The 10x libraries were sequenced on two alternative platforms (Illumina and Ultima Genomics). Twelve libraries were simultaneously sequenced on both platforms for comparison purposes and to demonstrate the scalability of our approach. We observed the Ultima-sequenced data to be highly similar to the Illumina-sequenced data (Extended Data Fig. 5a).

DNA production and sequencing

All healthy and patient DNA was produced from PB at sampling. DNA sequencing was performed on two targeted panels: the first a rich myeloid CH panel (InfiniSeq Myeloid Malignancies Panel, Sequentify, Israel) covering all known pre-leukemic mutations10 and the second a genotyping panel specifically designed to capture polymorphic sites prevalently expressed by RNA molecules from all cell types in our data. This allowed demultiplexing of individual pools based on individual specific SNP combinations and replaced previous, antibody-based multiplexing methods. Three to six individual samples were pooled on each experiment day after extraction of DNA aliquots, such that CD34 enrichment was performed on the entire pool of PBMCs produced. As with other methods of sample multiplexing, genotype-based multiplexing allows for robust doublet detection during data analysis, which enabled loading of 30,000–40,000 cells on each Chromium Chip lane.

Both our CH and genotyping panels are Molecular Inversion Probe (MIP) panels described in detail previously10. For the healthy cohort we used our CH panel v.3, containing 705 probes, covering leukemia-related SNVs and insertions/deletions (indels) in 47 genes, complemented by two amplicon-sequencing reactions to cover GC-rich regions in SRSF2 and ASXL1. For the cytopenic cohort, we used our CH panel v.4 (Supplementary Table 10). For alignment of reads we used Burrow–Wheeler Aligner (BWA)-MEM v.2 (ref. 33). As MIP sequencing is cost-effective yet noisy, we developed an in-house variant calling method to reliably identify low-VAF CH events10. For the genotyping panel we used Varscan for variant calling34. Each DNA sample was sequenced twice with a minimum depth of 106 paired-end reads on an Illumina Novaseq machine. Variant calling was performed as previously reported10. Our genotyping panel allows for the simultaneous detection of >2,000 SNVs. It includes heterozygous sites with at least a 5% minor allele frequency from the 1,000 Genomes project, which were extensively covered in our data (at least 80 unique molecular modifiers (UMIs) across all cells in a test 10x library), excluding sites in repetitive elements and sex chromosomes. Both panels were designed using MIPgen35 to ensure capture uniformity and specificity.

CH sequencing of high-RDW samples and controls

To compare propensity for CH and high-risk CH mutations in high-RDW cases and normal RDW controls, we performed deep targeted sequencing of DNA samples from 602 high-RDW (>15%) individuals, whose blood count did not meet MDS criteria (11.5 g dl−1 ≤ Hg ≤ 15.5 g dl−1 (F), 13 g dl−1 ≤ Hg ≤ 17 g dl−1 (M), 80 fL ≤ MCV ≤ 96 fl, PLT ≥ 100 × 109 l−1, Abs Neut ≥ 1.8 × 109 l−1) and 602 normal RDW, age- and sex-matched controls. Case–control matching was performed using the R MatchIt package, balanced on age and sex, method = ‘nearest’, ratio = 1, from a total of 18,147 individuals with longitudinal blood counts and available DNA at the Tel Aviv Sourasky Medical Center (TASMC) Integrated Cancer Prevention Center. All DNA samples were collected after obtaining written informed consent in accordance with the Declaration of Helsinki and were received de-identified from the TASMC. All relevant ethical regulations were followed and all protocols were approved by the TASMC ethics committee (under IRB protocol no. 02-130).

ScRNA-seq processing

We processed fastq files by executing cell-ranger (v.3.1.0) with an hg-38 reference genome. We filtered out cells with at least 20% mitochondrial expression, then removed mitochondrial genes (as well as few other batch-prone genes) and further filtered cells with ≤500 UMIs.

Doublet calling

We performed several steps to assign cells to individuals and to detect doublets. Our pipeline included the following steps:

    +
  1. + (1) + +

    Demultiplexing cells and calling doublets based on SNPs found in the scRNA-seq data

    + +
    2. +
  2. + (2) + +

    Building a metacell model using cells from all libraries, including cells previously marked as doublets, identifying and removing metacells made of doublets

    + +
    3. +
  3. + (3) + +

    Identifying doublet metacells based on expression of marker genes

    + +
    4. +
  4. + (4) + +

    Building the final metacell model and marking metacells as doublets based on expression markers

    + +
    5. +

In the first step, we identified doublets and assigned cells to individuals using Vireo v.0.3.2 (ref. 36) and Souporcell v.2.4 (ref. 37), which cluster cells based on SNPs found in sequenced RNA molecules. We executed Vireo (preceded by running cellsnp v.0.3.0) and Souporcell on each library separately. Both methods used SNPs from our genotyping panel10 which were covered by at least 20 UMIs in the library (in Souporcell—at least 10 from the major and minor allele each). We observed high agreement in doublet calling between the two methods.

In the next step, we built a metacell model with cells from all libraries. This model included cells that we already identified as doublets. The model was built with metacell2 (ref. 11), with a target metacell size of 200 cells. We then marked all metacells, where at least 35% of the cells were already marked as doublets, and all metacells that expressed key markers of distinct cell types as doublet metacells. All cells that belonged to a doublet metacell were then marked as doublets. We then built an additional metacell model (see below), without cells that were marked as doublets.

Assignment of cells to individuals

Vireo and Souporcell both cluster cells based on SNPs found in the sequenced RNA, such that cells in the same cluster belong to the same individual. We next assigned clusters of cells to the individual to whom they belonged. To this end, we correlated the genotypes of each cell cluster, as inferred by Vireo, to all genotypes that we measured using the MIP panel (only using sites with sufficient sequencing depth). As a control, this matching was performed against the MIP genotypes of all individuals in the cohort and not only those from the specific library analyzed. We observed clear matchings between Vireo clusters and individuals from the expected libraries in all cases. This method also correctly identified related individuals. The sex of all matched individuals was confirmed by expression of XIST in the RNA data.

Removal of droplets with megakaryocyte signatures

Droplets with complete or partial megakaryocyte expression (at least 5% of UMIs coming from a megakaryocyte gene program including PF4, PPBP and 131 additional genes) were removed from our model as a result of their overall high doublet rate, and a final metacell model was constructed from the retained cells ((1) not marked as doublets, (2) confidently assigned to an individual and (3) not exhibiting megakaryocyte expression).

Correcting for sequencing-platform bias

Our 10x libraries were sequenced on Ultima Genomics and Illumina sequencers. To minimize batch effects related to these sequencing-platform variations, we used libraries that were sequenced on both platforms to calculate an Illumina–Ultima correction factor per gene as the mean log2(fold-change) in expression of the gene across re-sequenced libraries. We then normalized each Ultima-sequenced library by downsampling genes with at least 0.28 log2(fold) Ultima overexpression and resampling genes with at least 0.2 Illumina overexpression. The downsampling and resampling were performed for each gene independently, across all cells in each Ultima library. The thresholds for downsampling and resampling were chosen such that the overall number of UMIs per cells remained similar; 87 genes with at least 4-fold-change between Ultima and Illumina were excluded from further processing.

Computing the reference metacell model

Our metacell model was built using metacell2, with a target metacell size of 200 cells, deriving 4,253 metacells. We marked histone, cell-cycle, ribosomal, sex-linked and stress response genes (including FOS and JUN) as forbidden genes, as well as genes with high technical variation, such as those with high or inconsistent differences between Illumina- and Ultima-sequenced technical replicates. These genes were not used for calculating gene–gene similarities but were included in downstream analyses. We annotated metacells using known markers as illustrated in Extended Data Fig. 1c. We excluded metacells with low CD34 expression, such as mature monocytes, B cells, T cells, natural killer (NK) cells, dendritic cells (DCs) and endothelial cells, as well as 20 GMP-L metacells, from most downstream analyses. We used UMAP projections of the metacell expression vector over genes with specific enrichment over cell types for visualization of the metacell manifold.

BM comparisons and projections

We used three BM datasets for comparison purposes: a dataset including CD34-enriched cells from two individual BM samples collected by us and processed similarly to PB (Fig. 1a), the Human Cell Atlas (HCA) BM dataset14 and a CD34+CD38− bead-based enriched BM dataset12. We previously processed and annotated the HCA dataset in a metacell model. We further constructed a metacell model for the two CD34-enriched BM samples collected by us using metacell2, in a similar fashion to that described previously, and downloaded the Setty et al. sequencing data12, processed it by running cell-ranger, and created a third BM metacell model from their data. To project our own PB data, our own BM data and the Setty CD34-enriched BM data on the HCA model, we correlated projected metacells from each of these models with HCA metacells over genes showing high variance in the HCA model. We annotated each Setty metacell using the mode of its five most correlated HCA metacells. We annotated our own BM data using both the mode of the five most correlated HCA metacells and expression of gene markers. We projected metacells from each of these models on the HCA UMAP using the mean x and y values of the five most correlated HCA metacells. To compare S-phase genes between PB and BM (Extended Data Fig. 3b), we calculated the S-phase signature (mean expression of six cell-cycle genes: TYMS, H2AFZ, PCNA, MCM4, HELLS and MKI67) for each PB and HCA metacell and plotted the distribution of these scores across metacells for each cell type.

HSC differentiation gene programs

To visualize transcriptional dynamics in HSCs, we sorted MEBEMP and CLP metacells based on their AVP expression. To calculate differential expression (DE) between HSCs and neighboring cell types (Extended Data Fig. 3g), we took the geometric mean expression of each gene across each of these cell states (within HSC or CLP-M or MEBEMP-E metacells) and calculated the difference of means between HSC and MEBEMP-E and between HSC and CLP-M metacells.

DE between individuals unexplained by the metacell model

We compared each individual’s pooled expression profile to a matched expression profile based on the individual’s distribution across metacells. We performed this analysis separately for MPPs or MEBEMPs (MPP, GMP-E, MEBEMP-E/-L, ERYP and BEMP) and CLPs (CLP-E/-M/-L and NKTDP). In each of these cell states, we downsampled each cell to have 500 UMIs and summed the UMIs across all cells of each individual, normalized the sum to 1 and calculated the log2(value) to obtain the observed expression. To compute matched expression, we downsampled each metacell to have 90,000 UMIs and summed all UMIs of the metacell to which each cell belongs for each individual. We normalized this matched expression to sum to 1 and took the log2(value). For Extended Data Fig. 3c,d, we plotted all genes that were expressed in either the observed or the matched expression in at least one individual (log2(expression) > 2−14.5 for MPPs or MEBEMPs, > 2−13.5 for CLPs), with at least a twofold change between matched and observed in at least one individual. We excluded genes exhibiting strong batch effects.

HSPC compositional analysis

To explore variance in cell-type composition between individuals, we first calculated the distribution of each individual’s cells across the CD34+ cell states. We further partitioned cells from the CD34+ cell states into finer-grained bins, using one HSC, four CLPs and ten MPP or MEBEMP bins, for a total of fifteen bins. We assigned HSC cells to bin 0, CLP-E cells to CLP bin 1 and CLP-M/-L cells to CLP bins 2–4 based on an AVP expression gradient, such that each of these bins consisted of an equal number of cells. We similarly assigned MPP and MEBEMP-E/-L cells into equal size MPP or MEBEMP bins 1–10 based on decreasing AVP expression.

The bottom panel of Fig. 2d shows individual enrichment across bins (log2(ratio of each individual’s cell frequency in each bin to the median cell frequency in that bin across individuals)). We partitioned individuals into three groups based on their mean enrichment across CLP bins 2–4—those with mean enrichment >0.5 are high CLP, those with <−0.5 are low and the rest are intermediate. We next defined the stemness score as the ratio between the number of cells in MPP or MEBEMP bins 1–5 and the total MPP or MEBEMP cells (bins 1–10). Individuals with stemness score >0.5 had enriched stemness. Individuals within each cluster were further sorted based on their stemness score. The combination of CLP enrichment and stemness defines the six classes shown in the figure.

Test for association between cell-state compositions and a numerical label

We used permutation tests to test for an association between cell-state distributions and a label, such as CBC indices or sync-scores. We sorted CD34+ cell states into 11 bins from late MEBEMP differentiation through HSCs to late CLP differentiation (as ordered in Fig. 2b). We correlated each of the 11 cell-state frequency vectors to the numerical label vector. We then looked at triplets of adjacent cell states in this order and calculated the mean correlation for each triplet to obtain nine mean correlation values and took the maximal absolute correlation value as a test statistic. We repeated this process after permuting the label vector 10,000× and used the test statistics from the permutations to derive a P value.

CD34+ cell frequency in the OneK1K dataset

We built a metacell model for the cells from Yazar et al.25. We labeled all cells in metacells with high CD34 expression (log2(fraction of UMIs > −14.3)) as CD34+ cells. We then selected individuals with at least 800 cells in the model and randomly sampled 800 cells from each (Supplementary Table 14). To produce Fig. 3d, we pooled these sampled cells by the decade of their individuals’ age and calculated the fraction of CD34+ cells in each decade. The 95% confidence intervals shown in the figure assume a binomial distribution, given the very sparse nature of the data.

Variably expressed gene modules

We detected gene modules with high variance across individuals while controlling for compositional variation. This was performed separately for myeloid and lymphoid states, in the following manner:

    +
  1. + (1) + +

    For each individual, we calculated the 5th percentile of their number of UMIs across all MPP metacell cells and downsampled all cells to this number. We then pooled all downsampled cells, normalized to sum to 1 and took the log2(value). This gave us the observed expression profile of each individual.

    + +
    2. +
  2. + (2) + +

    We then created the expected expression profile for each individual as follows: we partitioned all MPP metacells into 30 equal size bins based on their AVP expression, and downsampled metacells to 90,000 UMIs. We then took the average expression of each gene across downsampled metacells in each bin. This defined an expression profile for each of the 30 bins. To obtain an individual’s expected expression, we calculated the weighted average expression profile of bins, where the weight of each bin is proportional to the fraction of the individual’s cells from that bin, normalized to sum to 1 and took the log2(value). We then calculated the difference between the observed and expected expression profiles.

    + +
    3. +
  3. + (3) + +

    Our data showed some batch effect distinguishing samples collected in two calendric periods. As this effect could introduce co-variation between genes across individuals, we applied a correction controlling for it. This was performed using a linear model fitting each gene to the sample collection period. We then subtracted the inferred period factor from the samples that were collected in the second period. We found that this approach substantially reduced emergence of gene clusters linked with sample collection date bias.

    + +
    4. +
  4. + (4) + +

    We screened for genes with high variance that were unlikely to be affected residually by the main manifold differentiation process. We removed genes with high batch effects, genes with high AVP correlation (absolute value Pearson’s correlation >0.65), and genes highly correlated (absolute value Pearson’s correlation >0.5) with a module of differentially expressed genes between the first and second collection periods. We then calculated each gene’s variance in the difference between observed and expected expression across individuals. As some of this variance can be explained by sampling noise, we plotted each gene’s variance across individuals against its mean expression across individuals. We sorted genes by this expression value and subtracted from the variance of each gene a rolling mean of the variances of 100 neighboring genes in that ordering. We chose genes with variance at least 0.08 higher than the rolling mean variance.

    + +
    5. +
  5. + (5) + +

    We calculated a gene–gene Spearman’s correlation matrix for the high variance genes and clustered correlation profiles using hierarchical clustering. We removed genes with low mean correlation (<0.2) to their cluster and then removed gene clusters with low mean correlation between their genes (≤0.25 mean correlation for all gene pairs). We further computed gene–gene correlations using only samples from our first library collection period and required gene clusters to have a high mean correlation (>0.25) between their genes when using only these samples. We removed additional gene modules arising from this analysis resulting from batch effects or traces of MEBEMP differentiation not normalized by this approach. This resulted in Fig. 3f.

    + +
    6. +

We performed a similar analysis for CLPs, with few differences. The analysis included all cells from CLP-M metacells. The cells were partitioned into six equal size bins and partitioning was based on the average of their DNTT and VPREB1 expression. Genes with high absolute correlation to the average DNTT and VPREB1 expression were excluded. This was followed by hierarchical clustering of the gene–gene correlation profiles and removal of genes as described for MPPs.

Age regression

We developed age-regression models for MPP and CLP expression separately. To predict age, we used the difference between an individual’s observed and expected gene expression as described above (‘Variably expressed gene modules’). We used genes with minimal expression ≥2−14.5 for MPPs and ≥2−15.5 for CLPs across individuals. We trained a LASSO (least absolute shrinkage and selection operator) model using nested leave-one-out cross-validation. For each left-out sample, we performed cross-validation on the remaining samples to select LASSO’s λ parameter, trained a model using the selected λ and made a prediction on the left-out sample.

The LMNA signature

We used the difference between an individual’s observed and expected gene expression and correlated this difference to the difference in LMNA expression separately for MPPs and CLPs. We then summed the MPP and CLP correlation values and kept genes with summed correlation >0.7. We further removed genes with high technical variance, retaining 17 genes in the LMNA signature. To calculate individual LMNA signatures, we took the average value of these 17 genes in the observed–expected matrix of each individual for MPPs and CLPs separately. To plot Extended Data Fig. 6g, we calculated the geometric mean of LMNA signature gene expression for each individual in each one of the ten MPP or MEBEMP bins described earlier in Fig. 2d.

GoT analysis

GoT21 performed on sample N122 allowed us to mark this individual’s cells as wild-type or mutated. As a result of the low VAF of N122’s DNMT3A mutation, and to increase power, we marked cells with a DNMT3A mutation status that could not be determined by GoT as wild-type cells. For Extended Data Fig. 6b, we examined sample N122’s cell distribution across cell states.

Sync-score

We defined the AVP signature to include genes with high correlation (>0.6) to AVP across HSC, MPP and MEBEMP metacells and the GATA1 signature to include those with high correlation (>0.7) to GATA1. We filtered out genes with mean relative expression >2−10 in these metacells, to preclude a small number of genes from dominating the signatures. We then scored all HSC, MPP, MEBEMP-E and MEBEMP-L cells by their fraction of UMIs from the AVP and GATA1 signatures and partitioned them into 20 equal size bins of AVP signature expression and 20 equal size bins of GATA1 signature expression. The sync-score is then defined as the fraction of cells in GATA1 bins 13 and above (upper two quintiles of GATA1) that are in AVP bins 9 and above (upper three quintiles of AVP expression). To visualize sync-scores (Fig. 3i), we normalized this 20 × 20 bin matrix to sum to 1, smoothed the obtained matrix by averaging cells using a running window of length 3 and took the log2(value).

Differential gene expression with respect to age and CBC

DE was performed separately for MPP and CLP cells as well as for men and women. The MPP and CLP-M matrices previously used to detect variant gene modules were used here as well. Individual gene expression levels were correlated with age, maximal VAF of CH mutations and 20 CBC indices using Spearman’s correlation; the correlation was then tested for significance. The P values were false discovery rate (FDR)-corrected (Benjamini–Hochberg) for each label separately. For maximal VAF we additionally performed a Mann–Whitney U-test comparing individuals with and without detected mutations. DE between men and women was performed using a Mann–Whitney U-test on the same expression matrices.

Reconstruction of MDS classification models using improved cell mapping and filtering

For analyzing MDS classification we re-analyzed sequenced libraries of all disease cases and healthy individuals in two groups, separated by sequencing platform (Illumina and Ultima). Cell filtering was then applied for each of the two datasets using the process described above with the following minor modifications:

    +
  • +

    Re-mapping all cells using cell-ranger v.7.0.1

    +
    • +
  • +

    Both Vireo and Souporcell were limited to 7.4 M SNPs with minor allele frequency >0.05 according to the 1,000 Genome project, rather than SNPs from our genotyping panel

    +
    • +
  • +

    Refined filter for cell exclusion, excluding 10× particles (cells) with high mitochondrial content (>20%), platelet signature (PPBP > 0.2%), neutrophil signature (LCN2, CAMP and LTF) or erythrocyte signature (HBB, HBA1 and HBA2), and also excluding cells with low signature of nuclear RNAs (Supplementary Table 9 includes number of excluded cells)

    +
    • +
  • +

    Adjusting the doublet detection algorithm described above with an additional filter involving clustering cells and removal of cells with UMI count that is higher than 2.5-fold of their computed cluster median (thereby compensating for variable cell sizes across types). In addition, cells with high expression of both monocyte and MEBEMP markers were filtered out. Such extra steps were needed because, in some disease batches, highly specific cell states could contaminate other samples more than in standard reference batches.

    +
    • +

In silico sorting for inferring sample cHSPC compositions

To estimate cHSPC state for a given single-cell transcriptional profile, an in silico sorting scheme was developed (Extended Data Fig. 7a). First, our original reference model was used to compile gene signatures. Each signature was based on genes differentially expressed in a given cell type, such that the total number of UMIs for the signature is sufficiently high to allow classification at single-cell resolution (selected gene sets in Supplementary Table 13). Extended Data Fig. 7a shows the gating strategy used for classification using signature scores (log2(total signature UMI in cell) − log2(total cell UMI)). Cells with ambiguous gating were defined as unassigned. We confirmed that the gating strategy yields classification that is consistent with the annotation derived by applying metacell analysis to the new Fig. 4 reference model (see below) by projecting inferred classes on the metacell model UMAP projection (Extended Data Fig. 7b). We noted that sorting may reduce the manifold resolution compared with metacell analysis, but it provides robust results for downstream MDS classification purposes.

Healthy donor reference model for MDS analysis

Samples from 79 individuals who were sequenced on the Illumina platform and showed no evidence for disease were considered as the reference model for MDS classification (Fig. 4). This cohort was used for defining the normal distribution of cHSPC composition. It was also the basis for constructing a reference metacell model used for projecting patient data. This model includes 287,000 cells and 2,090 metacells, constructed using metacell2 with a target metacell size of 140 cells and other parameters similar to the original normal reference model.

Grouping MDS and cytopenia patients by composition

We used the inferred cHSPC compositions of 70 donors in the Fig. 4 reference model to score each composition abnormality of patients with MDS and cytopenia. Composition vectors p over cell types were log-transformed first as \({\mathbf{lp}}=\log_2({\epsilon +\mathbf{p}})\) where ϵ = 0.02. The distance between two samples was then defined using a Euclidean distance between their lp vectors. The abnormality score of a new sample was defined as the average distance for the four nearest neighbors in the reference model. The 0.98 quantile of the abnormality score of healthy donors (excluding reference donors) was used as a threshold (Extended Data Fig. 8a) for classifying patient compositions (excluding four patients exhibiting complex karyotypes) as normal or abnormal (Fig. 4b; GRP1 or GRP2-4, respectively). Patients with abnormal compositions were further grouped using hierarchical clustering of their lp vectors. Patients with normal compositions were ordered along a CLP frequency gradient in Fig. 4b (left), in analogy with Fig. 2d.

Estimating patients’ CNAs using scRNA projection on the reference

We constructed a metacell model from the filtered cells of each patient separately. This model was projected over the Fig. 4 reference model, using MCProj38. The result was a set of metacells for the patient, such that each metacell m was defined by its observed gene expression \({e}_{{{\mathrm{gm}}}}^{{{\mathrm{obs}}}}\) and projected gene expression \({e}_{{{\mathrm{gm}}}}^{{{\mathrm{proj}}}}\), as determined by MCProj using best matching reference behavior. Expression values were calculated using the geometric mean. Genes were filtered to remove sex-specific and sequencing-platform-specific genes (559 genes overall; Supplementary Table 15). Further filtering was done for 107 genes showing a consistent difference between observed and projected values over all individuals, as well as 27,317 lowly expressed genes (Supplementary Table 15).

To correct for batch effects leading to small GC content preference per library, we grouped genes into ten equal size bins according to the average GC content of 3′-scRNA-sequenced tags in representative sequenced libraries. For each gene g, we computed total observed and expected UMI counts, given the model’s projection on the reference:

$${n}_{g}^{{{\mathrm{obs}}}}=\sum _{c}{n}_{c}\times {e}_{{{\mathrm{gm}}}(c)}^{{{\mathrm{obs}}}}$$
$${n}_{g}^{{{\mathrm{proj}}}}=\sum _{c}{n}_{c}\times {e}_{{{\mathrm{gm}}}(c)}^{{{\mathrm{proj}}}}$$

where \({n}_{c}\) is the total number of UMIs for the cell c and \(m(c)\) is its metacell. The bias per GC bin \({{\mathrm{bias}}}_{{{\mathrm{GCbin}}}}\) is now approximately defined as the median of \(\frac{{n}_{g}^{{{\mathrm{obs}}}}}{{n}_{g}^{{{\mathrm{proj}}}}}\) across genes in the GC bin. In practice, we calculated the ratio \(\frac{{n}_{g}^{{{\mathrm{obs}}}}}{{n}_{g}^{{{\mathrm{proj}}}}}\) after normalizing \({n}_{g}^{{{\mathrm{obs}}}}\) and \({n}_{g}^{{proj}}\) by \({\sum }_{g{\prime} }{n}_{g{\prime} }^{{{\mathrm{obs}}}}\) and \({\sum }_{g{\prime} }{n}_{g{\prime} }^{{{\mathrm{proj}}}}\), respectively, and adding a regularization term of \({10}^{-5}\). We corrected each observed gene expression value \({e}_{{{\mathrm{gm}}}}^{{{\mathrm{obs}}}}\) by dividing by the appropriate \({{\mathrm{bias}}}_{{{\mathrm{GCbin}}}}\) value, generating \({e}_{{{\mathrm{gm}}}}^{{{\mathrm{ob}}}{s}^{{\prime} }}\), which is used to obtain a per-gene and metacell observed:expected expression log(ratio) \({\delta }_{{{\mathrm{gm}}}}=\log_2\left(\epsilon +{e}_{{{\mathrm{gm}}}}^{{{\mathrm{ob}}}{s}^{{\prime} }}\right)-\log_2\left(\epsilon +{e}_{{{\mathrm{gm}}}}^{{{\mathrm{proj}}}}\right)\), whereas \(\epsilon ={10}^{-5}\).

We split each chromosome to contiguous bins encompassing 20 genes (ignoring filtered genes). For each chromosome bin bchrom and each metacell m, the median log ratio was computed: \({\delta }_{m,{{\mathrm{bchrom}}}}={{\mathrm{Med}}}_{g\in {{\mathrm{bchrom}}}}({\delta }_{{{\mathrm{gm}}}})\).

The matrix \({\delta }_{m,{{\mathrm{bchrom}}}}\) of metacells and 20-gene bins describing estimated DE was then normalized by subtracting the median of each row (metacell) and visualized in a heatmap, where metacells were re-ordered using hierarchical clustering. Heatmaps of the derived matrices (see, for example, Extended Data Fig. 10a,b) were examined manually to identify CNAs (Extended Data Fig. 7c and Supplementary Table 12).

Within-state gene signatures

Within-state correlated gene sets were inferred from the reference metacell model by clustering the gene–metacell expression matrix of each annotated cell type, while considering only highly variable genes within the cell type. Clusters were evaluated and selected manually and expanded by adding correlated genes, resulting in the final gene sets (Supplementary Table 13). An MHC-II gene set was added after observation of MDS DE compared with the reference. The signature score per cell was estimated as the log-transformed normalized total UMI count for each gene set.

Signature scores per patient were extracted using the median signature of cells within a respective cell type (following the in silico sorting process described above). In the case of too few such cells, the signature score was considered missing (NA).

After classification of patients with cytopenia and MDS into groups 1–4 (Fig. 4b), we performed, for each within-state gene signature (and each relevant state), a Kruskal–Wallis test comparing signature expression levels between groups 1–4 and healthy donors. The signatures with the lowest P values were the MEBEMP-L MHC-II, S-phase and BEMP early signatures, which were accordingly shown in Fig. 4e and Extended Data Fig. 8d.

Features for MDS classification

The following features were collected to facilitate MDS classification:

    +
  • +

    CBC values: we used the values with minimal time gap from the cHSPC sampling date

    +
    • +
  • +

    Maximal CH VAF across mutations detected in the same blood sample that was used for scRNA-seq

    +
    • +
  • +

    The cHSPC compositions as inferred through in silico sorting

    +
    • +
  • +

    Twenty-one signature scores

    +
    • +
  • +

    Composition abnormality score (ʽGrouping MDS and cytopenia patients by compositionʼ)

    +
    • +
  • +

    Number of CNAs.

    +
    • +

We noted that signature scores might be missing (as a result of insufficient number of cells). In addition, a few individuals were missing CBC values.

All feature values per scRNA sample are included in Supplementary Table 9.

MDS classifier training and testing

XGBoost (xgboost Python package, v.2.0.3) training was performed on cohort 1 including 89 samples (41 normal, 20 cytopenia and 28 MDS). All of the samples in cohort 1 were sequenced on the Ultima platform. MDS (including MDS or myeloproliferative neoplasms) samples were considered positive, whereas cytopenia and normal samples were considered negative.

We applied feature selection separately in each leave-one-out fold, selecting a subset of the features for which the FDR-corrected Mann–Whitney P value for MDS or cytopenia separation was <0.1. We then inferred a classifier using the binary:logistic objective and default hyperparameters, except for tree_method = hist. Model performance was evaluated by pooling accuracy on the left-out samples of their respective folds.

Further evaluation of the approach was done by classifying cohort 2 (16 MDS and 9 cytopenia cases, all sequenced on the Illumina platform) using the classifier trained on cohort 1.

Blasts versus CLP-E analysis

We compared blast fraction from BM samples acquired at most 1 year before or after cHSPC sampling. Comparison was also restricted to cHSPC samples with >500 HSPCs (scRNA samples used in this analysis are specified in Supplementary Table 9). CLP-E-like frequency estimation was as described in Extended Data Fig. 9d.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

+