Resource

Clonally resolved single-cell multi-omics identifies

routes of cellular differentiation in acute myeloid
leukemia
Graphical abstract Authors
Sergi Beneyto-Calabuig,
Anne Kathrin Merbach,
Jonas-Alexander Kniffka, ...,
€ ller-Tidow,
Simon Raffel, Carsten Mu
Lars Velten

Correspondence
carsten.mueller-tidow@
med.uni-heidelberg.de (C.M.-T.),
[email protected] (L.V.)

In brief
Velten and colleagues develop
CloneTracer, a computational method to
identify clones in single-cell RNA-seq
data. Applied to immature cells from 19
acute myeloid leukemia patients,
CloneTracer shows that dormant
hematopoietic stem cells (HSCs) are
healthy or preleukemic. Leukemic stem
cells resemble healthy active HSCs but
give rise to aberrant progenitors.
Highlights
d CloneTracer extracts clonal information from single-cell
RNA-seq data

d Data resource of healthy and leukemic stem and progenitor

cells from 19 AML cases

d Dormant hematopoietic stem cells in AML patients are

healthy or preleukemic

d Leukemic stem cells resemble active HSCs but form aberrant

myeloid progenitors

Beneyto-Calabuig et al., 2023, Cell Stem Cell 30, 706–721

May 4, 2023 ª 2023 The Authors. Published by Elsevier Inc.
https://doi.org/10.1016/j.stem.2023.04.001 ll
 ll
OPEN ACCESS

Resource
Clonally resolved single-cell multi-omics identifies
routes of cellular differentiation
in acute myeloid leukemia
Sergi Beneyto-Calabuig,1,3,15 Anne Kathrin Merbach,2,4,15 Jonas-Alexander Kniffka,2,16 Magdalena Antes,2,5,6,16
Chelsea Szu-Tu,1,16 Christian Rohde,2,4 Alexander Waclawiczek,5,6 Patrick Stelmach,2,6 Sarah Graۧle,7,8,9 Philip Pervan,1
Maike Janssen, Jonathan J.M. Landry, Vladimir Benes, Anna Jauch, Michaela Brough, Marcus Bauer,12
2,4 10 10 11 11

Birgit Besenbeck,2 Julia Felden,2 Sebastian Ba €umer,13 Michael Hundemer,2 Tim Sauer,2 Caroline Pabst,2,4
12
Claudia Wickenhauser, Linus Angenendt, 13,14 Christoph Schliemann,13 Andreas Trumpp,5,6 Simon Haas,5,6,7,8,9
Michael Scherer,1 Simon Raffel,2 Carsten Mu € ller-Tidow,2,4,* and Lars Velten1,3,17,*
1Centre for Genomic Regulation (CRG), The Barcelona Institute of Science and Technology, Dr. Aiguader 88, Barcelona 08003, Spain
2Department of Medicine, Hematology, Oncology and Rheumatology, University Hospital Heidelberg, 69120 Heidelberg, Germany
3Universitat Pompeu Fabra (UPF), Barcelona, Spain
4Molecular Medicine Partnership Unit, European Molecular Biology Laboratory (EMBL), University of Heidelberg, 69117 Heidelberg, Germany
5Heidelberg Institute for Stem Cell Technology and Experimental Medicine (HI-STEM gGmbH), 69120 Heidelberg, Germany
6Division of Stem Cells and Cancer, Deutsches Krebsforschungszentrum (DKFZ) and DKFZ-ZMBH Alliance, 69120 Heidelberg, Germany

€tsmedizin Berlin, 10117 Berlin, Germany

7Berlin Institute of Health (BIH) at Charité – Universita
8Charité-Universita€tsmedizin, 10117 Berlin, Germany
9Berlin Institute for Medical Systems Biology, Max Delbru €ck Center for Molecular Medicine in the Helmholtz Association, 10115 Berlin,
Germany
10Genomics Core Facility, European Molecular Biology Laboratory (EMBL), 69117 Heidelberg, Germany
11Institute of Human Genetics, University of Heidelberg, 69120 Heidelberg, Germany
12Institute of Pathology, University Hospital Halle (Saale), Martin-Luther-University Halle-Wittenberg, 06112 Halle, Germany
13Department of Medicine A, Hematology and Oncology, University Hospital, Muenster, Germany
14Department of Biosystems Science and Engineering, ETH Zurich, Basel, Switzerland
15These authors contributed equally
16These authors contributed equally
17Lead contact

*Correspondence: [email protected] (C.M.-T.), [email protected] (L.V.)

https://doi.org/10.1016/j.stem.2023.04.001

SUMMARY

Inter-patient variability and the similarity of healthy and leukemic stem cells (LSCs) have impeded the char-
acterization of LSCs in acute myeloid leukemia (AML) and their differentiation landscape. Here, we introduce
CloneTracer, a novel method that adds clonal resolution to single-cell RNA-seq datasets. Applied to samples
from 19 AML patients, CloneTracer revealed routes of leukemic differentiation. Although residual healthy and
preleukemic cells dominated the dormant stem cell compartment, active LSCs resembled their healthy coun-
terpart and retained erythroid capacity. By contrast, downstream myeloid progenitors constituted a highly
aberrant, disease-defining compartment: their gene expression and differentiation state affected both the
chemotherapy response and leukemia’s ability to differentiate into transcriptomically normal monocytes.
Finally, we demonstrated the potential of CloneTracer to identify surface markers misregulated specifically
in leukemic cells. Taken together, CloneTracer reveals a differentiation landscape that mimics its healthy
counterpart and may determine biology and therapy response in AML.

INTRODUCTION thought to drive hematopoiesis, such as common myeloid pro-

genitors (CMPs), consist of mixtures of fully lineage committed
Our understanding of blood formation has fundamentally cells.2 Rather than passing through a CMP stage, lineage differ-
changed in the last decade. Single-cell RNA sequencing entiation occurs along two major branches, a lymphomyeloid
(scRNA-seq)-based studies have demonstrated that hematopoi- and an erythromyeloid branch.3,4 These results are supported
etic stem cells acquire priming early, at phenotypically immature by various functional assays.5–7 By contrast, the aberrations
stages.1 The oligopotent progenitor types that were previously that characterize the differentiation landscape in myeloid

706 Cell Stem Cell 30, 706–721, May 4, 2023 ª 2023 The Authors. Published by Elsevier Inc.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
 ll
Resource OPEN ACCESS

A B C

E F

G H I J

Figure 1. CloneTracer and Optimized 10x enable clonal tracking in droplet-based scRNA-seq
(A) Scheme of Optimized 10x.
(B) Normalized coverage across the mitochondrial genome obtained by default and Optimized 10x.
(legend continued on next page)
Cell Stem Cell 30, 706–721, May 4, 2023 707
 ll
OPEN ACCESS Resource

malignancies remain unknown. In particular, many diseases RESULTS

were thought to affect, or originate from, CMPs, which do not
represent a defined cell type. Improving coverage of nuclear and mitochondrial SNVs
Since the healthy and diseased hematopoietic systems in droplet-based scRNA-seq
co-exist in myeloid malignancies, investigating malignant differ- Conventional droplet-based scRNA-seq protocols exhibit low
entiation landscapes requires clonally resolved scRNA-seq coverage of nuclear and mitochondrial mutations. We mitigated
methods. Recent studies have profiled CALR-mutant8 this issue by splitting the cDNA pool after amplification and con-
and JAK2-mutant9 myeloproliferative neoplasm, as well as structing sequencing libraries specifically covering RNA expres-
DNMT3A-mutant clonal hematopoiesis10,11 and revealed the sion, surface antigen expression, nuclear SNVs, and mitochon-
expansion of particular differentiation states at the expense of drial genomes (Figure 1A; STAR Methods). In particular, we
others.12 However, the shape of the cellular differentiation land- constructed libraries that unlike default 10x Genomics cover
scape in full-blown acute myeloid leukemia (AML) remains the mitochondrial genome full length, similar to a recent report14
unknown: Are healthy routes of lineage differentiation co-opted (Figures 1B and S1A–S1C). In addition, 55%–85% of these li-
in this disease, or are novel, aberrant cellular identities created? braries mapped to the mitochondrial genome, allowing for a
In addition, more specifically, are leukemic stem cells (LSCs) a cost-effective deep sequencing of mitochondrial genomes.
consistent cell type resembling healthy HSCs, or are they hetero- False positive observations of mitochondrial genetic variants
geneous groups of leukemic cells that possess stemness prop- occurred at negligible background rates (Figures S1D–S1G).
erties? Answering these questions is of key importance to For genotyping of nuclear SNVs, we utilized a modified version
prioritize cellular targets for therapies and identify novel prog- of TAP-seq (targeted Perturb-seq)18 with nested, mutation-spe-
nostic factors. cific primers (Figures S1H–S1J and https://github.com/veltenlab/
Here, we have developed a new strategy to add clonal reso- CloneTracer/tree/master/primer_design for primer design soft-
lution to high-throughput (droplet-based) scRNA-seq data that ware that also assists with identifying mutations suitable for ampli-
robustly work across many of the heterogeneous AML geno- fication). Thereby, coverage on relevant mutations was substan-
types. Existing approaches use single nucleotide variants tially improved, compared with default 10x Genomics 30
(SNVs) or mitochondrial SNVs (mtSNVs) as qualitative markers (Figures 1C and S1K) and similar to results from related ap-
to identify healthy and malignant cells from scRNA-seq proaches.8,9,19 Mitochondrial and nuclear SNV-targeted
data.8–10,13,14 However, these measurements are noisy, and sequencing libraries can be constructed from existing (e.g.,
methods for quantitative analyses are lacking. Copy-number already sequenced) full-length cDNA libraries from 10x Genomics
variants (CNVs) can be inferred from scRNA-seq data15–17 but 30 , making this method (‘‘Optimized 10x’’) applicable to charac-
are not always present. Our new computational method, terize existing samples in more depth. Sequencing depth require-
CloneTracer, integrates information from SNVs, mtSNVs, and ments, as well as a comparison of long-read20 and short-read
infers CNVs (when present) through a statistical model appro- sequencing, are presented in the Methods S1. Libraries included
priate for noisy single-cell data. CloneTracer thereby identifies in the final dataset were analyzed with short-read sequencing.
clonal hierarchies and probabilistically assigns single cells to We compared the performance of Optimized 10x with a plate-
clones. based RNA-seq protocol (MutaSeq, a modified Smart-Seq2
We applied CloneTracer to bone marrow samples from 19 protocol21) and a droplet-based assay for transposase-acces-
AML patients. We showed that CloneTracer could unanimously sible chromatin with sequencing (ATAC-seq) method focused
identify most healthy and leukemic cells in 14 of these patients. on tracking mitochondrial mutations (sc-mito-ATAC-seq22)
By integrating data across all patients, we identified a population (Figures S2A–S2C). Optimized 10x maximized the mutational in-
of HSCs expressing a dormancy gene signature that was domi- formation available from default 10x Genomics libraries, and, un-
nated by residual healthy and preleukemic cells, as well as like low-throughput high-confidence plate-based methods,21,23
leukemic cells resembling active HSCs (active leukemic stem it displayed the throughput required for ambitious scRNA-seq
cells [aLSCs]) that often retained erythroid capacity. Down- oncology projects (Figure S2C).
stream of aLSCs, differentiation-blocked, aberrant myeloid
progenitors affected chemotherapy responses and fed into qual- CloneTracer, a statistical model to infer clonal
itatively normal myeloid differentiation. Together, our data estab- hierarchies and identities from scRNA-seq data
lished a healthy-like differentiation landscape that may deter- Despite the improved coverage of leukemic point mutations and
mine biology and therapy response in leukemia. mtSNVs in Optimized 10x, data from scRNA-seq-based

(C) Coverage of nuclear mutations from various AML patients. Only immature and early myeloid cells are included. See also Figure S1K.
(D) Illustration of the statistical challenge addressed by CloneTracer.
(E) Overview of two AML cohorts, see also Methods S1.
(F) Overview of longitudinal sampling in cohort B. Pie charts indicate clinical blast counts. DA/VA indicate treatment with Daunorubicin/Ara-C and Venetoclax/
Azacitidine, respectively.
(G) Top row: inferred clonal hierarchy for patient A.8. Middle row: stacked bar chart illustrating each cell’s probability to derive from the different clones shown in
top panel. Bottom rows: heatmap depicting the variant allele frequency of all clonal markers in all cells.
(H) Like (G), except that data from patient A.6 is shown.
(I) Clonal hierarchy of patient A.6 identified from sequencing of single-cell-derived colonies, see STAR Methods.
(J) Like (G), except that data from patient B.2 is shown. For CNVs, the scaled number of counts on the specified chromosome are shown.

708 Cell Stem Cell 30, 706–721, May 4, 2023

 ll
Resource OPEN ACCESS

protocols are inherently noisy, as illustrated by frequent allelic Here, a nuclear SNV with a high allele frequency in bulk
dropout even of highly expressed genes such as NPM1 (see exome-seq was located in MPO, which was covered in 22.8%
Figures 1C and S1K). For a confident interpretation of the data of cells. The mutant MPO allele was only observed in cells car-
and quantitative analyses, statistical methods are needed that rying a mitochondrial mutation (3019G>C) (Figure 1H). The mito-
identify the most likely hierarchy among the mutations and chondrial mutation was a suitable clonal marker, as it had likely
thereby, for example, clarify whether a mitochondrial mutation occurred before the nuclear mutation or there were only cells
or CNV is present in all cancer cells or demarcates a sub-clone. carrying both mutations. To verify these results, we grew sin-
Furthermore, dropout and false positive rates (FPRs) need to be gle-cell-derived colonies and genotyped MPO and the mito-
systematically accounted for when assigning cells to (sub-) chondrial mutations, confirming that the mitochondrial mutation
clones. is a high-confidence clonal leukemia marker (Figure 1I). Similar to
We therefore developed CloneTracer, a Bayesian model that patient A.6, mitochondrial mutations drove the assignment of
identifies the hierarchical relationship between mutations and leukemic and healthy cells in patient B.3.
assigns the cells to the clones. Our model considers previous In the case of patient B.2, we observed several sub-clonal
information, such as allele frequencies from exome sequencing mitochondrial mutations downstream of the co-occurring IDH2
(exome-seq), and most importantly, it accounts for the tech- and DNMT3A mutations that, when occurring together, likely
nical noise associated with single-cell measurements of constitute a leukemic, and not preleukemic, event24 (Figure 1J).
CNVs, SNVs, and mtSNVs (Figure 1D; see Methods S1 for Since these genes displayed high dropout and no reliable clonal
detail). Thereby, it first compares possible clonal hierarchies. marker was identified, there was often considerable uncertainty
Second, for the mutational hierarchy with the highest evidence, regarding the assignment of healthy vs. leukemic cells. Hence,
it computes the posterior probability of each cell to belong to this patient was excluded from the clonal analysis. In the remain-
any particular clone. ing 4 patients, there were no well-covered clonal markers (i.e.,
We applied CloneTracer to 19 AML patients from two cohorts mtSNVs, CNVs, or well-covered SNVs).
(Figures 1E and 1F; Table S3). Cohort A consisted of diagnostic Together, these examples illustrate the importance of using
bone marrow samples that were subjected to a CD34 enrich- statistical models when interpreting single-cell genotyping
ment before single-cell CITE-seq (cellular indexing of transcrip- data. All subsequent analyses involving clonal identities pertain
tomes and epitopes by sequencing); a median of 2,232 single to the 14 patients with high-confidence healthy/leukemic assign-
cells per patient passed stringent quality control filters. Somatic ments. Importantly, ‘‘leukemic’’ is here defined purely by the
variants were identified a priori by ATAC-seq (for mitochondrial presence of a mutation not observed in the T cell lineage (usually
variants) and exome-seq (for nuclear variants) of myeloid and NPM1 or a CNV, and in some cases, a mitochondrial marker mu-
T cells (Figure S1G; Tables S3 and S5). Cohort B included paired tation) and not by a functional ability to induce leukemia. Recent
longitudinal samples from four individual patients at the time of work has used clonal tracking to demonstrate that not all stem
diagnosis, after therapy, and (in one case) at the time of relapse. cells carrying a leukemic driver are functionally leukemogenic.25
A median of 12,034 single cells per patient passed quality filters.
Somatic variants were identified a priori by panel sequencing. In Validation of CloneTracer
both cohorts, cells were stained with CITE-seq surface anti- Overall, 91% of the cells from the 14 patients could be assigned
bodies (see Table S1). Overall, we analyzed 88,602 single cells as healthy or leukemic with high confidence (Figure 2A; STAR
from 25 specimens. Methods). We validated these assignments using established
To demonstrate the performance of CloneTracer, we chose to parameters for AML diagnosis: In AML, most myeloid cells are
highlight three representative patients (A.8, A.6, and B.2). leukemic, whereas lymphoid cells are usually healthy.26,27 We
Detailed analyses of all patients are described in Methods S1. identify lymphoid and myeloid cells from scRNA-seq data and
Patient A.8 represents the performance of CloneTracer in used this assignment as an indicator for healthy and leukemic,
cases with well-covered leukemic mutations: respectively. Under this assumption, the median area under
Here, a mutation in NPM1 was covered in 94% of the cells. the receiver-operating characteristics curve (AUROC) of
Mutations in RPS29 and a mitochondrial gene co-occurred CloneTracer was 0.96 (range 0.88–1). The discretized
with this mutation. A preleukemic DNMT3A mutation occurred CloneTracer assignments had a median FPR across patients of
upstream of NPM1 but displayed a higher dropout rate. 9% (range 0.1%–20%) and a median false negative rate (FNR)
CloneTracer confidently assigned cells as part of the leukemic of 1% (range 0%–20%, Figures 2B and 2C). For patients with
clone if any of the downstream (NPM1, RPS29, or the mtSNV) SNVs as clonal markers, statistically naive assignments8,28 that
mutations were observed (Figure 1G). In their absence, there classify a cell as leukemic if at least one mutant allele is observed
was often no conclusive evidence if the cell was healthy or pre- and otherwise as healthy if at least one healthy allele is observed,
leukemic, due to the dropout of DNMT3A. We grouped these sub-optimally balanced between FPR and FNR (Figure 2C).
cells as ‘healthy’ in downstream analyses and followed up on Notably, not all myeloid cells in AML patients are leukemic. We
preleukemia for selected patients (below, see Figures 4 and therefore considered leukemia-associated immunophenotypes
S5). Similar results were obtained for 11 further patients with a (LAIPs) to distinguish leukemic vs. healthy myeloid cells. LAIPs
well-covered mutation (mostly NPM1 or a CNV) on top of the are cell state-specific, aberrantly expressed markers identified
clonal hierarchy. during the routine clinical flow cytometric analysis of AML diag-
Patient A.6 represents the behavior of CloneTracer in cases nostic samples.29,30 Three patients carried a significant number
with moderately covered leukemic mutations and co-occurring of residual healthy cells along the full myeloid differentiation
mitochondrial markers: spectrum, as well as a clinically described LAIP. In these

Cell Stem Cell 30, 706–721, May 4, 2023 709

 ll
OPEN ACCESS Resource

A B C

D E F

Figure 2. Validation of CloneTracer clonal assignments

(A) Pie chart summarizing the assignment of cells as healthy or leukemic.
(B) Barchart depicting the fraction of cells assigned as healthy (blue), leukemic (red), or indeterminate (gray), stratified by cell type. Cell types with less than 50 cells
covered are excluded.
(C) ROC curves computed from CloneTracer leukemia probabilities, assuming that lymphoid cells are healthy and myeloid cells are leukemic. Dots depict
statistically naive point estimates for the patients without CNVs.
(D) Gene expression data were projected to a healthy reference,4 and for myeloid progenitors, a pseudotime was computed. Plots depict smoothened expression
of LAIP markers over pseudotime, stratified by the clone. Points indicate mean expression within 20 equally sized bins, and point size indicates number of cells
per bin. Asterisks indicate significance of differential expression. *** FDR < 0.001, ** FDR < 0.01, * FDR < 0.1. p values are from a Wilcoxon test of library-size
normalized ADT counts.
(E) As in (D), but for patient B.1.
(F) Scatterplot depicting p values from statistical tests comparing surface marker expression between healthy and leukemic immature cells. x axis shows es-
timates obtained using a statistically naive assignment, and y axis shows estimates obtained using CloneTracer assignments.

patients, we projected each cell to a pseudotime of myeloid dif- according to the known differentiation trajectories (Figures 3A
ferentiation (see also below) and plotted the protein expression and S3A, points in color). By contrast, cells from leukemic pa-
of clinically identified LAIP markers over differentiation pseudo- tients abundantly existed in cell states not observed in healthy
time separately for leukemic and healthy cells. As expected, patients (Figure 3A, gray points). Many of these cell states
leukemic, but not residual healthy cells, expressed LAIP in a dif- were observed in single or few patients, highlighting inter-patient
ferentiation state-dependent manner (Figures 2D and 2E). For heterogeneity (Figures 3B, 3C, and S3B).
example, CD7 was only expressed by leukemic stem-like cells Highlighting CloneTracer assignments on the uMAP showed
in patient B.1. The statistical power for identifying LAIP markers that most lymphoid cells from AML patients were healthy and
was increased by using CloneTracer, compared with the statis- myeloid cells were leukemic (Figure 3D; see also Figure 2B).
tically naive assignment (Figure 2F). Few lymphoid cells assigned as leukemic likely constituted
Together, these analyses demonstrate that CloneTracer false positive calls. By contrast, significant numbers of healthy
correctly assigned cells as healthy and leukemic and outper- monocytes and healthy HSC/MPP (multipotent progenitor)-like
formed statistically naive assignments. cells occurred in 6 and 5 patients, respectively (Figures 3D
and S3B).
Differentiation hierarchies in leukemia We next aimed to identify leukemic and healthy stem cells. We
To identify differentiation landscapes, we integrated gene observed a cluster, C6 (see Figure S4A for exact cluster labels),
expression data from all 19 patients with data from two healthy that contained HSCs from the healthy reference individuals (Fig-
individuals (A.0 from this study and C.3 from Triana et al.4). The ure 3A), as well as both healthy and leukemic cells from the
integration strategy was selected to preserve real biological dif- different AML patients (Figures 3B and 3D). Interestingly, C6
ferences between samples (Figure S2D) while accounting for contained cells from 16 of 19 AML patients, whereas most other
technical batch effects (Figure S2E; see STAR Methods). This progenitor clusters were dominated by cells from only one
analysis showed that cells from the healthy individuals arrange patient each (Figures 3B and 3C). Other progenitor populations

710 Cell Stem Cell 30, 706–721, May 4, 2023

 ll
Resource OPEN ACCESS

A B C

D E

F G H

Figure 3. Differentiation landscapes in AML

(A) Uniform manifold approximation and projection (uMAP) depicting integrated data from both cohorts. Color: cell type for cells from healthy individuals (A.0 and
C.34), see also Figure S3A. Gray: cells from leukemia patients.
(B) Same uMAP highlighting patient identity.
(C) Bar chart summarizing the number of patients represented in each cluster with at least 5 cells. Inset: histogram depicting the size of C6 as a fraction of total
bone marrow. n = 16 patients with cells in C6 are included.
(D) uMAP highlighting CloneTracer leukemia probabilities for 14/19 AML patients. Gray: cells from remaining individuals.
(E) Volcano plot highlighting the number of patients where a given surface marker was significantly (p < 0.05) differentially expressed between cells from C6 and
other immature myeloid cells from the same patient, vs. the average log2-fold change across patients. n = 16 patients with C6 represented and n = 2 healthy
individuals were analyzed.
(F) Volcano plot as in (E), but for RNA expression. To avoid overplotting, a score from 0 to 1 that depended on the log sum of p values was added to the patient
number on the y axis. 30Color: number of patients where the gene appeared as significant in label-retaining (LRC) vs. non-LRC AML cells.31

(legend continued on next page)

Cell Stem Cell 30, 706–721, May 4, 2023 711

 ll
OPEN ACCESS Resource

appeared to ‘‘emerge’’ from cluster C6, which was more evident To evaluate the potential leukemic and preleukemic content of
in a 3D uMAP (see https://veltenlab.crg.eu/clonetracer/ ). the dormant stem cell population, we increased the numbers of
Compared with other progenitor cells, cells from cluster C6 analyzed cells.
tended to express stem cell surface markers (Figure 3E: expres- First, we focused on 3 patients with NPM1 mutations where
sions of CD34, CD90, and CD49f and lower expressions of CD38 CD34 expression was rare and nearly exclusive to the dormant
and CD45RA). Interestingly, C6 overexpressed genes that were stem cell population (Figure 4C: patients A.10, A.11, and A.12).
identified as upregulated in label-retaining AML cells (LRCs) dur- We sorted CD34+ cells and performed MutaSeq,21 a well-based
ing xenotransplant assays, compared with the non-label-retain- single-cell method that allows us to efficiently capture mutations
ing fraction31 (Figure 3F; Table S2). LRCs were characterized as in lowly expressed genes such as DNMT3A. The sorting strategy
the population responsible for causing disease in xenotrans- resulted in a significant number of cells expressing the dormancy
plants and for drug resistance.31 Across the different AML pa- gene signature (Figure 4D). For all of these dormant stem cells,
tients, cluster C6 contained a median of 1% (range 0.1%– genotype data indicated that they were healthy or preleukemic
17%) of the total bone marrow (Figure 3C, inset), which is higher (Figure 4E).
than the LSC number estimated from xenotransplants.35 To follow up on dormant stem cells in two patients where
To evaluate the similarity of cells from C6 to healthy stem cells, CD34 expression was not exclusive to the dormant stem cell
we projected all cells to a healthy reference4 and assigned each compartment, we sequenced 23,110 additional CD34+ and total
cell to the most similar healthy cell state (see STAR Methods; BM cells from two patients, A.2 and A.9 (Figure S5A). In the case
Figure 3G) and a score that quantifies the similarity (Figure 3H). of A.9, we thereby identified 267 C6 cells, a subset of which ex-
Leukemic cells from cluster C6 were very similar to healthy pressed the dormancy signature (Figures S5B and S5D). All but
HSCs, whereas leukemic cells outside of C6 mapped to HSCs two C6 cells here were healthy or preleukemic (i.e., DNMT3A-
or downstream myeloid progenitor states but displayed lower mutant). In the case of A.2, we identified 1,109 cluster C6 cells
similarity. At the level of monocytes and dendritic cells, the tran- that were mostly leukemic and lacked the expression of the
scriptomic similarity between leukemic cells and healthy cells dormancy score (Figures S5B and S5C). These data further al-
increased (Figure 3H, inset). We therefore conceptually structure lowed us to demonstrate that the main dataset was sufficiently
leukemic differentiation in three stages, putative healthy-like powered to detect all cell states (Figures S5C and S5D).
stem cells (‘‘C6’’), highly heterogeneous and aberrant progeni- Taken together, our data suggest that the dormant stem cell
tors, and mature, healthy-like monocytes/dendritic cells. Impor- compartment is predominantly healthy or preleukemic. By
tantly, most leukemic cells in cluster C6 were marked by muta- contrast, the active stem cell compartment was leukemic in 12
tions in NPM1 or CNVs, which are typically associated with of the 14 patients. Our results cannot rule out the existence of
leukemia and not preleukemia27 (Figures S3C and S3D). rare leukemic dormant stem cells that might be relevant for
relapse.
The dormant stem cell compartment is healthy or
preleukemic LSCs retain erythroid capacity
To further characterize the putative stem cell cluster C6, we We next investigated the leukemic fraction of C6 and its routes of
focused on the six patients for whom both healthy and leukemic differentiation. In some patients, leukemic cells from C6 ex-
cells occurred within this cluster. Healthy and leukemic cells in pressed ‘‘active HSC’’ genes32,33 or ‘‘high-output HSC’’ genes36
C6 were generally separated by the principal component anal- relative to healthy cells from C6 (Figures 4B and S4B; Table S2).
ysis based on gene expression (Figure 4A). Healthy cells ex- To identify whether these cells are truly distinct from other
pressed genes characterized as ‘‘dormant HSC’’ genes in (1) a leukemic progenitors, we performed differential expression
recent scRNA-seq study of highly purified human HSCs,32 (2) la- testing, contrasting leukemic C6 cells to other leukemic myeloid
bel retention assays in mice,33 and (3) ‘‘low-output HSC’’ genes progenitor cells from the same patient. Although most leukemic
identified using clonal tracking36 (Figures 4B and S4B; Table S2). cells expressed genes associated with lymphomyeloid priming,
A dormant HSC gene expression signature robustly separated leukemic C6 cells highly expressed genes associated with early
healthy from the large majority of leukemic C6 cells across pa- erythromyeloid (erythroid, megakaryocytic, and eosinophilic/
tients (Figure 4C); cells expressing the dormant signature were basophilic) priming,1 AP-1 transcription factors, and genes asso-
consistently CD34+CD38 (Figure S4C). These results suggest ciated with LRCs31 (Figures 5A, S4D, and S4E; Table S2). In line
that in AML, the dormant stem cell compartment, where present with this observation, in 10 of the 14 patients with confident
or observed, contains healthy stem cells (dHSCs). By contrast, CloneTracer assignments, we observed erythroid progenitors
the active stem cell compartment was predominantly leukemic: carrying leukemic mutations (Figure 5B). Typically, these cells
these cells were predominantly healthy only in 2 of the 14 inves- carried NPM1 mutations and/or CNVs and were hence derived
tigated patients (A.9, A.13) (Figure 4C). Of note, five of the seven from the leukemic clone, and not from a preleukemic clone
patients where we observe dormant healthy HSCs belong to the (Figures S3C and S3D). The abundance of mutation-carrying
karyotypically normal, NPM1 mutant subtype (Figure 4C). erythroid progenitor cells correlated with the abundance of C6

(G) Cells from leukemia patients were projected to the healthy reference4. Color: most similar healthy cell type for each cell (STAR Methods). See Figure S3A for
color legend.
(H) uMAP highlighting the similarity score, i.e., similarity to the 5 most similar healthy reference cells (STAR Methods). Inset: smoothened average of the similarity
score over pseudotime.

712 Cell Stem Cell 30, 706–721, May 4, 2023

 ll
Resource OPEN ACCESS

A B

Figure 4. Characterization of LSCs

(A) PCA of RNA expression data of cells from C6 was performed separately for each patient. Score plots from n = 6 patients with both healthy and leukemic cells
represented in C6 are shown.
(legend continued on next page)
Cell Stem Cell 30, 706–721, May 4, 2023 713
 ll
OPEN ACCESS Resource

(Figure 5B). These results indicate that most leukemias, specif- ure S3B). The fraction of healthy monocytes correlated inversely
ically C6, can differentiate into the erythroid lineage at low rates. with the overall number of monocytes (Figure 5E): in leukemia
Together, these results allowed us to designate leukemic cells cells with a few monocytes (e.g., FAB [French-American-British
from C6 as aLSCs, a rare population of stem-like cells that exists AML classification] M0,M1), these monocytes were derived
in most AML patients and that often retain erythroid capacity. from residual healthy stem cells. An unsupervised analysis of
gene expressions revealed that monocytes exist in two cell
The state and extent of the differentiation block states. Leukemia-derived monocytes were enriched in a cell
determine the phenotypic manifestation of AML state with higher expressions of MHC-II and interferon response
We next focused on the leukemic progenitors downstream of genes (IFITM1 and IFITM3) (Figure 5F; Table S2). A similar signa-
LSCs. To determine the healthy cell state most strongly resem- ture was recently described for monocytes in clonal
bling these cells, their transcriptome was projected onto healthy hematopoiesis.38
progenitor cells, ranging from early MPPs to promyelocytes4 Together, these results suggest that leukemia-derived mono-
(see also Figures 3G and 3H). We thereby obtained an average cytes originated from incomplete differentiation blocks at the
pseudotime (i.e., a value describing each cell’s progression progenitor level and matured along normal differentiation path-
along the stem cell to monocyte trajectory). The average ways. The stage of the differentiation block at MPP, LMPP, or
projected pseudotime was associated with therapy response promyelocyte stages was linked to the first-line chemotherapy
(Figure 5C; only n = 14 patients treated with anthracycline and response. Furthermore, the strength of the differentiation block
cytarabine induction therapies were included here). We found was also encoded at the progenitor level and determined the de-
that patients with the most immature leukemic progenitors had gree of monocytic differentiation. The stage and the degree of
blast persistence or died during the first induction therapy, the differentiation block are independent properties.
whereas patients with LMPP (lymphomyeloid primed progeni- Our results suggested that the stage of the differentiation
tors)-like leukemic progenitors went into complete remission block was an important feature of the AML and thus may be
(p = 0.04, Wilcoxon test). Although the cohort size underlying subject to clonal selection. To investigate this hypothesis, we
these analyses was small, the results are consistent with a recent focused on three patients with co-existing sub-clones marked
report studying the relationship between differentiation arrest by relevant driver mutations (Figure S6). In these patients,
and survival in a large bulk RNA-seq cohort.37 The presence or sub-clones shifted toward more immature differentiation
size of the C6 stem cell population was not correlated with blocks, compared with the parental clones, possibly because
chemotherapy response. Taken together, these results suggest evolutionary pressures may favor differentiation blocks at
that the stage of the differentiation block may play a role in deter- more immature states. Given the small patient number available
mining chemotherapy response. for the analyses of sub-clones, we cannot rule out that other
We next investigated the ability of the leukemic progenitors to properties or genetic drift led to the expansion of the sub-
give rise to mature monocytes, which was highly variable be- clones.
tween patients. As expected, the genotype (e.g., NPM1 mutant)
only partly explained the degree of monocytic differentiation. We CloneTracer enables the discovery of leukemia and
hypothesized that leukemic progenitors with a larger resem- healthy specific markers
blance to their healthy equivalent might display a weaker differ- Our results indicated that LSCs, as well as leukemia-derived
entiation block. We computed for each progenitor cell a similarity monocytes, are rather healthy like and difficult to distinguish
score, describing how close it resembled the most similar from their healthy counterparts. This raised the question of
healthy cell (see Figure 3H). We found that this score, after ac- whether specific markers can be used to identify healthy vs.
counting for the genotype, correlated closely with the fraction leukemic cells of various differentiation stages, including stem
of mature monocytes or dendritic cells in the bone marrow (Fig- cells and monocytes. Such markers can possibly be identified
ure 5D). Patients with aberrant progenitor cells had a few mature by comparing healthy and leukemic cells from the same pa-
myeloid cells. By contrast, patients with progenitors more tient, thereby avoiding batch effects, genetic background,
closely resembling their healthy counterparts had large numbers and other variables typically confounding healthy-cancer com-
of mature cells. Taken together, these results suggest that the parisons. To investigate this idea, we first used the CITE-seq
‘‘degree’’ of the differentiation block, together with leukemia’s data of cohort B, since a larger number of surface markers
genotype, determines the fraction of monocytes in the bone and cells per patient were covered. We asked whether there
marrow. are markers that are overexpressed or depleted in leukemic
Of note, we observed patients with mostly healthy monocytes cells of various differentiation stages, compared with healthy
and other patients with mostly leukemic monocytes (see Fig- cells of the same stage. These comparisons identified CD11c

(B) Volcano plot as in Figure 3F, comparing healthy and leukemic cells from C6. n = 6 patients were analyzed, as in (A). Genes are colored by human HSC gene
signatures.32 phyper: hypergeometric test for enrichment.
(C) PCA of cells from cluster 6 performed jointly across all patients but using exclusively genes from the dHSC signature.32,33 Cells to the right of the dotted line are
putative dormant stem cells.
(D) Rare putatively healthy dormant CD34+ stem cells from three patients were sorted and subjected to a well-based scRNA-seq protocol.21 uMAP plots, from left
to right, show: (i) Projection on original uMAP from Figure 3, (ii) uMAP of Smart-Seq2 data, (iii) CD34 expression, (iv) HLF34 expression, and (v) a dormancy score
computed from the gene list in.32,33
(E) uMAPs as in (D), highlighting the variant allele frequency of relevant preleukemic and leukemic mutations.

714 Cell Stem Cell 30, 706–721, May 4, 2023

 ll
Resource OPEN ACCESS

A B

C D

E F

Figure 5. Differentiation pathways downstream of LSCs

(A) Volcano plot as in Figure 3F, comparing leukemic cells from C6 to leukemic immature myeloid cells from the same patient. n = 14 patients were analyzed.
Color: priming gene signatures.1
(B) Left panel: zoom-in on Figure 3D, displaying only clusters 6, 15, and 29. C15 is a patient-specific cluster displaying aberrant expression of hemoglobins. Right
panel: scatterplot depicting the size of cluster 6 and cluster 29 as a fraction of immature leukemic cells.
(C) Left panel: uMAP highlighting myeloid pseudotime obtained from projection to a healthy reference4. Right panel: boxplot contrasting the therapy response of
each patient with the average pseudotime of the patients’ immature leukemic cells.
(D) Scatterplot depicting the fraction of mature myeloid cells in diagnostic bone marrow samples across n=19 leukemic and one healthy individual (A.0) as a
function of the average similarity score of the immature myeloid progenitors, see also Figure 3H. Genotype is color coded. For patient B.4, the fraction of mature
myeloid cells was computed separately for the two sub-clones.
(E) Scatterplot depicting the fraction of monocytes in total bone marrow (x axis) and the fraction of monocytes that are healthy (y axis) in n = 14 patients.
(F) Left panel: CloneTracer assignments on the uMAP of cluster 1 and 2 (monocytes). Right panel: volcano plot as in Figure 3F, comparing cluster C1 to cluster C2
from the same patient across n = 19 patients and n = 2 healthy individuals.

Cell Stem Cell 30, 706–721, May 4, 2023 715

 ll
OPEN ACCESS Resource

A B C

D E

F G H

I J K L

Figure 6. Discovery of leukemia and healthy specific markers

(A) Scatterplots of the expression of CD49f and CD11c highlighting the clonal identity for patients B.1 and B.4. Only CD14- cells are shown. Numbers indicate the
percentage of leukemic cells in each of the four quadrants. p values are from a Fisher test for the association of quadrant with clonal identity.
(B) Like (A), except that only CD14+ cells are shown.
(C) Validation of CD11c and CD49f as leukemia/healthy markers by FACS sorting followed by FISH analysis. See STAR Methods. Gates were arbitrarily set to
achieve sufficient cell numbers. p values were calculated using a Fisher test. Right panel: representative images show hybridization of FISH probes.
(D) Schematic overview of the xenotransplantation experiments.
(E) Engraftment and lineage potential of CD34+CD14CD11c+CD49f (CG), CD34+CD14CD11cCD49f+ (HG), and CD34+CD14 peripheral blood and bone
marrow cells isolated from three de novo AML patients. Numbers indicate the quantity of mice with engraftment versus the total number of mice. Scale bars
indicate the standard deviation.
(F) Scheme illustrating the use of TMAs.
(G) TMA data from n = 86 patients. Ratio between the healthy and cancer gate was calculated as a function of FAB classification. p value was calculated using a
Wilcoxon test.

(legend continued on next page)

716 Cell Stem Cell 30, 706–721, May 4, 2023
 ll
Resource OPEN ACCESS

as overexpressed by leukemic cells and CD49f as enriched in particularly helpful in AML diagnosis if large numbers of mono-
healthy cells (Figure S7A). Accordingly, we observed an enrich- cytes are present.
ment of leukemic and healthy cells in the CD11c+CD49f and In stem cells, the CD11c+/CD49f combination was informa-
CD11cCD49f+ fraction, respectively (Figures 6A and 6B). tive in a subset of patients. In 5 of the 6 patients analyzed by
Since CD11c expression changed as a function of differentia- CloneTracer who contained healthy and leukemic cells in cluster
tion (Figure S7B), enrichment analyses of leukemic and healthy C6, CD49f was more highly expressed by healthy (i.e., dormant)
cells were performed for the immature (CD14) and mature stem cells40 (Figure 6K). The expression of CD11c on CD34+
(CD14+) compartments separately (Figures 6A and 6B). cells was variable across and within genotypes (Figures 6L and
We next confirmed the specificity of the CD11c/CD49f marker S7F). Accordingly, data integration by single-cell transcriptomics
combination by FACS sorting followed by fluorescent in situ hy- was overall superior in identifying the multipotent leukemia stem
bridization (FISH). In patient B.4, leukemic cells carrying the cell cluster, as well as stem-like progenitor cells, compared with
monosomy 7 were enriched in the CD11c+CD49f fraction, flow cytometry.
whereas healthy cells diploid for chromosome 7 were enriched
in the CD11cCD49f+ fraction (Figures 6C and S7C) (p = DISCUSSION
1010). Similar enrichments were demonstrated in an indepen-
dent patient with trisomy 8 (Figure S7D). To investigate routes of cellular differentiation in AML, we have
To further demonstrate that CD11c and CD49f can be used to introduced CloneTracer, a computational method for adding
enrich for functionally healthy cells, we performed xenotrans- clonal resolution and identifying leukemic and healthy cells in
plantation assays. We sorted CD34+CD14 peripheral blood scRNA-seq data. Tailored to scRNA-seq, CloneTracer extends
and bone marrow cells from three de novo AML patients into on DNA-seq-specific error models,40–42 as well as models that
CD11c+CD49f and CD11cCD49f+ fractions and trans- require previous knowledge of the clonal hierarchy.43 In the
planted each fraction into two immunocompromised mice. We AML context, CloneTracer confidently identified healthy and
observed that the putatively healthy CD11cCD49f+ fractions leukemic cells in 14/19 patients. CloneTracer assignments
gave rise to both myeloid and lymphoid engraftment in 8/10 relied on the presence of a clonal CNV (observed in 5 patients),
NSG mice from 3 of the 3 patients, indicating healthy hematopoi- a clonal mutation in a highly expressed nuclear gene (observed
esis (Figures 6D and 6E). The putatively leukemic CD11c+ in 7 patients), and/or a clonal mitochondrial mutation (observed
CD49f fractions did not engraft (Figures 6D and 6E), as is in 6 patients; full detail for all patients is provided in the
frequently observed for de novo AML samples.39 Methods S1). By combining the three layers of information,
Thus, the marker combination CD11c and CD49f identified CloneTracer outperformed methods that look at individual
CD34+ progenitor populations in AML specimens, which repo- layers only.8,14,17 Previous knowledge of the mutations is
pulate NSG mice with healthy cells. In the samples studied, the required to run CloneTracer, and we recommend calling these
marker combination could not be used to enrich rare LSCs. mutations from bulk data (e.g., exome sequencing, bulk ATAC-
Finally, we evaluated these markers in two larger cohorts with seq for mitochondrial variants, and karyotyping), although mito-
different techniques: chondrial variants and CNVs can also be called de novo from
We used immunohistochemistry for tissue microarrays (TMAs) single-cell data.14,17
from 86 AML patients and analyzed expression of CD14, CD34, The availability of clonal information enabled us to clarify routes
CD11c, and CD49f (Figures 6F–6H and S7E). A distinct cohort of of leukemic differentiation. Of note, through the integration of data
87 AML patients was analyzed by flow cytometry for expression from all patients, we identified a cluster of stem cells that con-
of CD14, CD34, and CD11c (Figures 6I and 6J). Together, these sisted of dormant, healthy, or preleukemic stem cells and pre-
results allowed us to provide a perspective on the specificity of dominantly leukemic, active SCs with retained erythroid potential
CD11c and CD49f and their potential relevance as markers for and a gene expression signature resembling label-retaining cells
healthy vs. leukemic cells. In particular, we observed that in in xenotransplants.31 Since both dHSCs/dpre (dormant preleuke-
more differentiated leukemias (FAB M2–M5), the putatively leu- mic cells)-LSCs and aLSCs exhibited relatively consistent gene
kemia-derived CD14+ cells were predominantly CD11c+ expression signatures across patients and healthy individuals,
CD49f. In undifferentiated leukemias (M0), which contain only data integration of scRNA-seq datasets represented a robust
a small number of monocytes, however, residual, putatively strategy for their identification. By contrast, these cells are difficult
healthy CD14+ cells showed a CD11cCD49f+ phenotype to enrich through flow sorting strategies: the large degree of inter-
(TMA data, Figures 6F–6H and S7E) and decreased CD11c patient heterogeneity at the level of progenitors renders it difficult
expression (Figures 6I and 6J). Thus, at the level of monocytes, to develop universal purification schemes.
CD11c and CD49f constituted a robust combination of markers Downstream of LSCs, we observed a highly heterogeneous
to quantitate the fraction of leukemia content. This might be and aberrant compartment of immature myeloid cells. The

(H) Representative mid-optical sections of a CD14, CD11c, and CD49f stained tissue micro array used for quantification in (G) (see also Figure S7E for scale bar).
Arrows, upper row: CD14+/CD11c/CD49f+ cell. Arrow, lower row: CD14+/CD11c+/CD49f cell.
(I) Scheme illustrating the flow cytometry experiment.
(J) Scatterplot relating the CD11c mean fluorescent intensity on CD14+ cells to the fraction of bone marrow that is CD14+, and the genotype. Flow cytometry data
from n = 59 individuals of selected genotypes is shown.
(K) Boxplot comparing the expression of CD49f in healthy and leukemic cells from cluster C6.
(L) Bar chart relating the expression of CD11c on CD34+ cells to genotype across n = 87 patients profiled by flow cytometry.

Cell Stem Cell 30, 706–721, May 4, 2023 717

 ll
OPEN ACCESS Resource

Figure 7. Model of leukemia differentiation

Adapted from a model of healthy hematopoietic differentiation.1

patient-specific stage of the differentiation arrest observed in cult to amplify and large chromosomal inversions or transloca-
this compartment determined the initial chemotherapy response tions in non-coding regions cannot be mapped. For DNMT3A,
of the patient. The strength of the block independently deter- coverage was obtained in approximately 20% of the CD34+
mined the overall degree of monocytic differentiation. Unlike cells; hence, preleukemic cells were difficult to distinguish from
cellular hierarchies identified from bulk data,37 the availability healthy cells with high confidence. Preleukemic cells were there-
of single-cell resolution allowed us to distinguish stem cells fore followed up on with well-based protocols (Figures 4D and
(C6) from various immature, ‘‘stem-like’’ progenitors and 4E) or simply by sequencing larger numbers of cells (Figure S5).
pinpoint a poor first-line chemotherapy response to the latter. In the future, methods that combine DNA-based genotyping45
Figure 7 summarizes our model of leukemic differentiation with RNA-seq in droplets might overcome the limitations of
pathways and suggests overall similarities to healthy hematopoi- CloneTracer but might initially suffer from worse quality of the
etic differentiation, but with an aberrant myeloid progenitor RNA-seq data.
compartment. Our model suggests that leukemic mutations, At the level of the cohort analyzed, a limitation of the study is
although present in stem cells and monocytes, may prominently that statements are drawn from only 19 patients (14 of whom
exert their effect in the cellular context of progenitors: Seven of have clonal tracking information). Hence, the results may not
the ten most commonly mutated AML driver genes44 were be entirely representative of the large heterogeneity of AML
more strongly expressed in progenitors, compared with stem genotypes and phenotypes observed, and studies with larger
cells and monocytes (Figure S7G). Differences in expression cohorts are going to systematically link genotypes and scRNA-
levels might lead to specific effects of the mutated gene in seq phenotypes.
each cellular compartment.
This implies that AML evolution requires mutations in slowly STAR+METHODS
dividing stem cells, although selection occurs at the level of pro-
genitors. Such a model would be in line with the low number of Detailed methods are provided in the online version of this paper
genetic aberrations observed in most AMLs. Of note, our data and include the following:
are static and cannot exclude the possibility that progenitor cells
in AML might also de-differentiate to give rise to stem cells. d KEY RESOURCES TABLE
In sum, these data may carry implications for the future devel- d RESOURCE AVAILABILITY
opment of therapeutic strategies: our results indicate that in B Lead contact
most if not all AML patients, there is a stem cell compartment B Materials availability
distinct from the most immature progenitor cells. Hence, tar- B Data and code availability
geted therapies aimed at immature progenitors may increase d EXPERIMENTAL MODEL AND SUBJECT DETAILS
the initial therapeutic response, but unless these therapies also B Human subjects
target the actual stem cell compartment, the effect on relapse B Animals
and long-term survival might be limited. d METHOD DETAILS
B Collection of bone marrow
Limitations of the study B Panel, exome, and bulk ATAC sequencing
At the level of the specific single-cell methodology employed for B Antibody-oligo conjugation
clonal tracking, a limitation is that in droplet-based scRNA-seq B CITEseq surface labeling and FACS sorting
protocols, SNVs in lowly expressed genes such as TET2 are diffi- B Single-cell RNA sequencing

718 Cell Stem Cell 30, 706–721, May 4, 2023

 ll
Resource OPEN ACCESS

B Optimized 10x: Mitochondrial libraries research. All other authors were involved in the acquisition and characteriza-
B Optimized 10x: Targeted genotyping libraries tion of clinical specimen. L.V., A.K.M., S.B.-C., and C.M.-T. wrote the manu-
B Plate-based single-cell RNA-seq (MutaSeq) script and generated the figures. All authors have read and commented on
the manuscript.
B Genotyping of single cell derived cultures
B Raw 10x Genomics data processing
DECLARATION OF INTERESTS
B Analysis of single cell gene expression data
B Dormancy score calculation The Department of Medicine V (Director C.M.-T.) receives research funding
B Analysis of DNAseq from single-cell derived colonies from multiple pharmaceutical and biotech companies especially for clinical tri-
als but also for translational research.
B Processing of MutaSeq scRNAseq data
B Raw data processing of MAESTER data
Received: July 29, 2022
B Fluorescent In Situ Hybridization Revised: February 5, 2023
B Tissue microarrays Accepted: March 30, 2023
B Large cohort flow cytometry analysis Published: April 24, 2023
B Xenotransplantations
d QUANTIFICATION AND STATISTICAL ANALYSIS REFERENCES
B CloneTracer model
1. Velten, L., Haas, S.F., Raffel, S., Blaszkiewicz, S., Islam, S., Hennig, B.P.,
B Differential expression testing
Hirche, C., Lutz, C., Buss, E.C., Nowak, D., et al. (2017). Human haemato-
B Data visualization poietic stem cell lineage commitment is a continuous process. Nat. Cell
d ADDITIONAL RESOURCES Biol. 19, 271–281. https://doi.org/10.1038/ncb3493.
2. Paul, F., Arkin, Y., Giladi, A., Jaitin, D., Kenigsberg, E., Keren-Shaul, H.,
SUPPLEMENTAL INFORMATION Winter, D., Lara-Astiaso, D., Gury, M., Weiner, A., et al. (2015).
Transcriptional heterogeneity and lineage commitment in myeloid progen-
Supplemental information can be found online at https://doi.org/10.1016/j. itors. Cell 163, 1663–1677. https://doi.org/10.1016/j.cell.2015.11.013.
stem.2023.04.001. 3. Tusi, B.K., Wolock, S.L., Weinreb, C., Hwang, Y., Hidalgo, D., Zilionis, R.,
Waisman, A., Huh, J.R., Klein, A.M., and Socolovsky, M. (2018).
ACKNOWLEDGMENTS Population snapshots predict early haematopoietic and erythroid hierar-
chies. Nature 555, 54–60. https://doi.org/10.1038/nature25741.
We thank Fengbiao Zhou, Anna Mathioudaki, and Judith Zaugg for discus- 4. Triana, S., Vonficht, D., Jopp-Saile, L., Raffel, S., Lutz, R., Leonce, D.,
sions and Laleh Haghverdi and Valerie Marot-Laussazaie for providing feed- Antes, M., Hernández-Malmierca, P., Ordoñez-Rueda, D., Ramasz, B.,
back on the mathematical model. We thank all members of GeneCore et al. (2021). Single-cell proteo-genomic reference maps of the hemato-
(EMBL) for assistance with the CITE-seq experiments, the DKFZ Single-Cell poietic system enable the purification and massive profiling of precisely
Open Lab (scOpenLab) for assistance with the MutaSeq/SmartSeq2 experi- defined cell states. Nat. Immunol. 22, 1577–1589. https://doi.org/10.
ment, the FACS facilities of DKFZ, Clinics HD, and CRG/UPF, and the geno- 1038/s41590-021-01059-0.
mics units at CRG and CNAG. We thank the NCT CLB, Sektion Cell Bio-
5. Perié, L., Duffy, K.R., Kok, L., de Boer, R.J., and Schumacher, T.N. (2015).
banking, for processing and providing bone marrow samples. Figure 1A was
The branching point in erythro-myeloid differentiation. Cell 163, 1655–
created with BioRender.com. Icons used in the graphical abstract are from
1662. https://doi.org/10.1016/j.cell.2015.11.059.
Servier Medical Art and licensed under CC-BY 3.0. This work was financially
supported by the German Bundesministerium fu €r Bildung und Forschung 6. Rodriguez-Fraticelli, A.E., Wolock, S.L., Weinreb, C.S., Panero, R., Patel,
(BMBF) through the Juniorverbund in der Systemmedizin ‘‘LeukoSyStem’’ S.H., Jankovic, M., Sun, J., Calogero, R.A., Klein, A.M., and Camargo, F.D.
(FKZ 01ZX1911D to L.V. and S.R.) as well as the Verbundprojekt SMART- (2018). Clonal analysis of lineage fate in native haematopoiesis. Nature
CARE (031L0212A to C.M.-T.), the Emerson foundation grant 643577 (to 553, 212–216. https://doi.org/10.1038/nature25168.
L.V.), grant PID2019-108082GA-I00 and PRE2020-093229 by the Spanish 7. Notta, F., Zandi, S., Takayama, N., Dobson, S., Gan, O.I., Wilson, G.,
Ministry of Science, Innovation and Universities (MCIU/AEI/FEDER, UE), the Kaufmann, K.B., McLeod, J., Laurenti, E., Dunant, C.F., et al. (2016).
German Research Foundation (DFG; projects MU1328/18-1 and MU1328/ Distinct routes of lineage development reshape the human blood hierarchy
21-1 and MU1328/23-1 to C.M.-T.), and the German Cancer Aid (DKH; project across ontogeny. Science 351, aab2116. https://doi.org/10.1126/science.
70113908 to C.M.-T.). L.V. acknowledges support of the Spanish Ministry of aab2116.
Science and Innovation to the EMBL partnership, the Centro de Excelencia 8. Nam, A.S., Kim, K.T., Chaligne, R., Izzo, F., Ang, C., Taylor, J., Myers,
Severo Ochoa and the CERCA Programme/Generalitat de Catalunya. R.M., Abu-Zeinah, G., Brand, R., Omans, N.D., et al. (2019). Somatic mu-
C.M.-T., A.K.M., and J.-A.K. gratefully acknowledge the data storage service tations and cell identity linked by Genotyping of transcriptomes. Nature
SDS@hd supported by the Ministry of Science, Research and the Arts Baden- 571, 355–360. https://doi.org/10.1038/s41586-019-1367-0.
Wu€rttemberg (MWK) and the German Research Foundation (DFG) through
9. van Egeren, D., Escabi, J., Nguyen, M., Liu, S., Reilly, C.R., Patel, S.,
grant INST 35/1314-1 FUGG and INST 35/1503-1 FUGG. J.-A.K. acknowl-
Kamaz, B., Kalyva, M., DeAngelo, D.J., Galinsky, I., et al. (2021).
edges support of the Deutsche Gesellschaft fu €r Ha€matologie und Medizini-
Reconstructing the lineage histories and differentiation trajectories of indi-
sche Onkologie e.V. (DGHO) and Deutsche José Carreras Leuka €mie-Stiftung
vidual cancer cells in myeloproliferative neoplasms. Cell Stem Cell 28,
e.V. through the José Carreras-DGHO-Promotionsstipendium.
514–523.e9. https://doi.org/10.1016/j.stem.2021.02.001.
10. Nam, A.S., Dusaj, N., Izzo, F., Murali, R., Mouhieddine, T.H., Myers, R.M.,
AUTHOR CONTRIBUTIONS
Sotelo, J., Benbarche, S., Gaiti, F., Tahri, S., et al. (2022). Single-cell multi-
A.K.M., S.R., C.M.-T., and L.V. conceived the project. C.S.-T., A.K.M., M.A., omics in human clonal hematopoiesis reveals that DNMT3A R882 muta-
J.-A.K., M.J., A.W., P.S., and M.B. generated the data and developed the lab- tions perturb early progenitor states through selective hypomethylation.
oratory protocols. S.B.-C., J.-A.K., A.K.M., M.B., P.P., and L.V. analyzed the https://doi.org/10.1101/2022.01.14.476225.
data with support from M.S. and C.R. S.B.-C. and L.V. developed the statisti- 11. Izzo, F., Lee, S.C., Poran, A., Chaligne, R., Gaiti, F., Gross, B., Murali, R.R.,
cal model. J.J.-M.L. and V.B. generated and processed raw sequencing data. Deochand, S.D., Ang, C., Jones, P.W., et al. (2020). DNA methylation
S.B. advised on antibody-oligo conjugation. A.W. performed xenotransplant disruption reshapes the hematopoietic differentiation landscape. Nat.
experiments. A.J. and M.B. performed FISH. L.V. and C.M.-T. supervised Genet. 52, 378–387. https://doi.org/10.1038/s41588-020-0595-4.

Cell Stem Cell 30, 706–721, May 4, 2023 719

 ll
OPEN ACCESS Resource
12. Nam, A.S., Chaligne, R., and Landau, D.A. (2021). Integrating genetic and 27. Shlush, L.I., Zandi, S., Mitchell, A., Chen, W.C., Brandwein, J.M., Gupta,
non-genetic determinants of cancer evolution by single-cell multi-omics. V., Kennedy, J.A., Schimmer, A.D., Schuh, A.C., Yee, K.W., et al. (2014).
Nat. Rev. Genet. 22, 3–18. https://doi.org/10.1038/s41576-020-0265-5. Identification of pre-leukaemic haematopoietic stem cells in acute
13. Petti, A.A., Williams, S.R., Miller, C.A., Fiddes, I.T., Srivatsan, S.N., Chen, leukaemia. Nature 506, 328–333. https://doi.org/10.1038/nature13038.
D.Y., Fronick, C.C., Fulton, R.S., Church, D.M., and Ley, T.J. (2019). A gen- 28. van Galen, P., Hovestadt, V., Wadsworth, M.H., Hughes, T.K., Griffin, G.K.,
eral approach for detecting expressed mutations in AML cells using single Battaglia, S., Verga, J.A., Stephansky, J., Pastika, T.J., Lombardi Story, J.,
cell RNA-sequencing. Nat. Commun. 10, 3660. https://doi.org/10.1038/ et al. (2019). Single-cell RNA-seq reveals AML hierarchies relevant to dis-
s41467-019-11591-1. ease progression and immunity. Cell 176, 1265–1281.e24. https://doi.org/
14. Miller, T.E., Lareau, C.A., Verga, J.A., DePasquale, E.A.K., Liu, V., Ssozi, 10.1016/j.cell.2019.01.031.
D., Sandor, K., Yin, Y., Ludwig, L.S., el Farran, C.A., et al. (2022). 29. Feller, N., van der Velden, V.H.J., Brooimans, R.A., Boeckx, N., Preijers, F.,
Mitochondrial variant enrichment from high-throughput single-cell RNA Kelder, A., de Greef, I., Westra, G., te Marvelde, J.G., Aerts, P., et al.
sequencing resolves clonal populations. Nat. Biotechnol. 40, 1030– (2013). Defining consensus leukemia-associated immunophenotypes for
1034. https://doi.org/10.1038/s41587-022-01210-8. detection of minimal residual disease in acute myeloid leukemia in a multi-
15. Patel, A.P., Tirosh, I., Trombetta, J.J., Shalek, A.K., Gillespie, S.M., center setting. Blood Cancer J. 3, e129. https://doi.org/10.1038/bcj.
Wakimoto, H., Cahill, D.P., Nahed, B.V., Curry, W.T., Martuza, R.L., 2013.27.
et al. (2014). Single-cell RNA-seq highlights intratumoral heterogeneity in 30. Al-Mawali, A., Gillis, D., Hissaria, P., and Lewis, I. (2008). Incidence,
primary glioblastoma. Science 344, 1396–1401. https://doi.org/10.1126/ sensitivity, and specificity of leukemia-associated phenotypes in
science.1254257. acute myeloid leukemia using specific five-color multiparameter flow
16. Gao, R., Bai, S., Henderson, Y.C., Lin, Y., Schalck, A., Yan, Y., Kumar, T., cytometry. Am. J. Clin. Pathol. 129, 934–945. https://doi.org/10.1309/
Hu, M., Sei, E., Davis, A., et al. (2021). Delineating copy number and clonal FY0UMAMM91VPMR2W.
substructure in human tumors from single-cell transcriptomes. Nat. 31. Takao, S., Morell, V., Brown, F.C., Koche, R., and Kentsis, A. (2022).
Biotechnol. 39, 599–608. https://doi.org/10.1038/s41587-020-00795-2. Epigenetic mechanisms controlling human leukemia stem cells and ther-
17. Gao, T., Soldatov, R., Sarkar, H., Kurkiewicz, A., Biederstedt, E., Loh, P.R., apy resistance. https://doi.org/10.1101/2022.09.22.509005.
and Kharchenko, P.V. (2023). Haplotype-aware analysis of somatic copy 32. Zhang, Y.W., Mess, J., Aizarani, N., Mishra, P., Johnson, C., Romero-
number variations from single-cell transcriptomes. Nat. Biotechnol. 41, €cklein, K.,
Mulero, M.C., Rettkowski, J., Schönberger, K., Obier, N., Ja
417–426. https://doi.org/10.1038/s41587-022-01468-y. et al. (2022). Hyaluronic Acid–GPRC5C Signalling Promotes Dormancy
18. Schraivogel, D., Gschwind, A.R., Milbank, J.H., Leonce, D.R., Jakob, P., in Haematopoietic Stem Cells (Springer). https://doi.org/10.1038/
Mathur, L., Korbel, J.O., Merten, C.A., Velten, L., and Steinmetz, L.M. s41556-022-00931-x.
(2020). Targeted Perturb-seq enables genome-scale genetic screens in 33. Cabezas-Wallscheid, N., Buettner, F., Sommerkamp, P., Klimmeck, D.,
single cells. Nat. Methods 17, 629–635. https://doi.org/10.1038/s41592- Ladel, L., Thalheimer, F.B., Pastor-Flores, D., Roma, L.P., Renders, S.,
020-0837-5. Zeisberger, P., et al. (2017). Vitamin A-retinoic acid signaling regulates he-
19. Gohl, D.M., Magli, A., Garbe, J., Becker, A., Johnson, D.M., Anderson, S., matopoietic stem cell dormancy. Cell 169, 807–823.e19. https://doi.org/
Auch, B., Billstein, B., Froehling, E., McDevitt, S.L., et al. (2019). Measuring 10.1016/j.cell.2017.04.018.
sequencer size bias using REcount: a novel method for highly accurate 34. Lehnertz, B., Chagraoui, J., MacRae, T., Tomellini, E., Corneau, S.,
Illumina sequencing-based quantification. Genome Biol. 20, 85. https:// Mayotte, N., Boivin, I., Durand, A., Gracias, D., and Sauvageau, G.
doi.org/10.1186/s13059-019-1691-6. (2021). HLF expression defines the human hematopoietic stem cell state.
20. Lebrigand, K., Magnone, V., Barbry, P., and Waldmann, R. (2020). High Blood 138, 2642–2654. https://doi.org/10.1182/blood.2021010745.
throughput error corrected nanopore single cell transcriptome 35. Bonnet, D., and Dick, J.E. (1997). Human acute myeloid leukemia is orga-
sequencing. Nat. Commun. 11, 4025. https://doi.org/10.1038/s41467- nized as a hierarchy that originates from a primitive hematopoietic cell.
020-17800-6. Nat. Med. 3, 730–737. https://doi.org/10.1038/nm0797-730.
21. Velten, L., Story, B.A., Hernández-Malmierca, P., Raffel, S., Leonce, D.R., 36. Rodriguez-Fraticelli, A.E., Weinreb, C., Wang, S.W., Migueles, R.P.,
Milbank, J., Paulsen, M., Demir, A., Szu-Tu, C., Frömel, R., et al. (2021). Jankovic, M., Usart, M., Klein, A.M., Lowell, S., and Camargo, F.D.
Identification of leukemic and pre-leukemic stem cells by clonal tracking (2020). Single-cell lineage tracing unveils a role for TCF15 in haematopoi-
from single-cell transcriptomics. Nat. Commun. 12, 1366. https://doi. esis. Nature 583, 585–589. https://doi.org/10.1038/s41586-020-2503-6.
org/10.1038/s41467-021-21650-1. 37. Zeng, A.G.X., Bansal, S., Jin, L., Mitchell, A., Chen, W.C., Abbas, H.A.,
22. Lareau, C.A., Ludwig, L.S., Muus, C., Gohil, S.H., Zhao, T., Chiang, Z., Pelka, Chan-Seng-Yue, M., Voisin, V., van Galen, P., Tierens, A., et al. (2022). A
K., Verboon, J.M., Luo, W., Christian, E., et al. (2021). Massively parallel sin- cellular hierarchy framework for understanding heterogeneity and predict-
gle-cell mitochondrial DNA genotyping and chromatin profiling. Nat. ing drug response in acute myeloid leukemia. Nat. Med. 28, 1212–1223.
Biotechnol. 39, 451–461. https://doi.org/10.1038/s41587-020-0645-6. https://doi.org/10.1038/s41591-022-01819-x.
23. Rodriguez-Meira, A., Buck, G., Clark, S.A., Povinelli, B.J., Alcolea, V., 38. Brett Heimlich, J., Bhat, P., Parker, A.C., Jenkins, M.T., Vlasschaert, C.,
Louka, E., McGowan, S., Hamblin, A., Sousos, N., Barkas, N., et al. Ulloa, J., Potts, C.R., Olson, S., Silver, A.J., Ahmad, A., et al. (2022).
(2019). Unravelling intratumoral heterogeneity through high-sensitivity sin- Mutated cells mediate distinct inflammatory responses in clonal hemato-
gle-cell mutational analysis and parallel RNA sequencing. Mol. Cell 73, poiesis. https://doi.org/10.1101/2022.12.01.518580.
1292–1305.e8. https://doi.org/10.1016/j.molcel.2019.01.009. 39. Krevvata, M., Shan, X., Zhou, C., dos Santos, C., Habineza Ndikuyeze, G.,
24. Zhang, X., Wang, X., Wang, X.Q.D., Su, J., Putluri, N., Zhou, T., Qu, Y., Secreto, A., Glover, J., Trotman, W., Brake-Silla, G., Nunez-Cruz, S., et al.
Jeong, M., Guzman, A., Rosas, C., et al. (2020). Dnmt3a loss and Idh2 neo- (2018). Cytokines increase engraftment of human acute myeloid leukemia
morphic mutations mutually potentiate malignant hematopoiesis. Blood cells in immunocompromised mice but not engraftment of human myelo-
135, 845–856. https://doi.org/10.1182/blood.2019003330. dysplastic syndrome cells. Haematologica 103, 959–971. https://doi.org/
25. Fennell, K.A., Vassiliadis, D., Lam, E.Y.N., Martelotto, L.G., Balic, J.J., 10.3324/haematol.2017.183202.
Hollizeck, S., Weber, T.S., Semple, T., Wang, Q., Miles, D.C., et al. 40. Jahn, K., Kuipers, J., and Beerenwinkel, N. (2016). Tree inference for sin-
(2022). Non-genetic determinants of malignant clonal fitness at single- gle-cell data. Genome Biol. 17, 86. https://doi.org/10.1186/s13059-016-
cell resolution. Nature 601, 125–131. https://doi.org/10.1038/s41586- 0936-x.
021-04206-7. 41. Malikic, S., Mehrabadi, F.R., Ciccolella, S., Rahman, M.K., Ricketts, C.,
26. Pelcovits, A., and Niroula, R. (2020). Acute myeloid leukemia: a review. R. I. Haghshenas, E., Seidman, D., Hach, F., Hajirasouliha, I., and Sahinalp,
Med. J. (2013) 103, 38–40. S.C. (2019). PhISCS: a combinatorial approach for subperfect tumor

720 Cell Stem Cell 30, 706–721, May 4, 2023

 ll
Resource OPEN ACCESS

phylogeny reconstruction via integrative use of single-cell and bulk 50. Macosko, E.Z., Basu, A., Satija, R., Nemesh, J., Shekhar, K., Goldman, M.,
sequencing data. Genome Res. 29, 1860–1877. https://doi.org/10.1101/ Tirosh, I., Bialas, A.R., Kamitaki, N., Martersteck, E.M., et al. (2015). Highly
gr.234435.118. parallel genome-wide expression profiling of individual cells using nanoli-
42. Zafar, H., Navin, N., Chen, K., and Nakhleh, L. (2019). SiCloneFit: Bayesian ter droplets. Cell 161, 1202–1214. https://doi.org/10.1016/j.cell.2015.
inference of population structure, genotype, and phylogeny of tumor 05.002.
clones from single-cell genome sequencing data. Genome Res. 29, 51. Mölder, F., Jablonski, K.P., Letcher, B., Hall, M.B., Tomkins-Tinch, C.H.,
1847–1859. https://doi.org/10.1101/gr.243121.118. Sochat, V., Forster, J., Lee, S., Twardziok, S.O., Kanitz, A., et al. (2021).
43. McCarthy, D.J., Rostom, R., Huang, Y., Kunz, D.J., Danecek, P., Bonder, Sustainable data analysis with Snakemake. F1000Res 10, 33. https://
M.J., Hagai, T., Lyu, R.; HipSci Consortium, and Wang, W., et al. (2020). doi.org/10.12688/f1000research.29032.2.
Cardelino: computational integration of somatic clonal substructure and 52. Kiselev, V.Y., Yiu, A., and Hemberg, M. (2018). scmap: projection of single-
single-cell transcriptomes. Nat. Methods 17, 414–421. https://doi.org/ cell RNA-seq data across data sets. Nat. Methods 15, 359–362. https://
10.1038/s41592-020-0766-3. doi.org/10.1038/nmeth.4644.
44. Papaemmanuil, E., Gerstung, M., Bullinger, L., Gaidzik, V.I., Paschka, P.,
53. Hie, B., Bryson, B., and Berger, B. (2019). Efficient integration of heteroge-
Roberts, N.D., Potter, N.E., Heuser, M., Thol, F., Bolli, N., et al. (2016).
neous single-cell transcriptomes using Scanorama. Nat. Biotechnol. 37,
Genomic classification and prognosis in acute myeloid leukemia.
685–691. https://doi.org/10.1038/s41587-019-0113-3.
N. Engl. J. Med. 374, 2209–2221. https://doi.org/10.1056/
NEJMoa1516192. 54. Stuart, T., Butler, A., Hoffman, P., Hafemeister, C., Papalexi, E., Mauck,
W.M., Hao, Y., Stoeckius, M., Smibert, P., and Satija, R. (2019).
45. Miles, L.A., Bowman, R.L., Merlinsky, T.R., Csete, I.S., Ooi, A.T., Durruthy-
Comprehensive integration of single-cell data. Cell 177, 1888–1902.e21.
Durruthy, R., Bowman, M., Famulare, C., Patel, M.A., Mendez, P., et al.
https://doi.org/10.1016/j.cell.2019.05.031.
(2020). Single-cell mutation analysis of clonal evolution in myeloid malig-
nancies. Nature 587, 477–482. https://doi.org/10.1038/s41586-020- 55. Luecken, M.D., Bu €ttner, M., Chaichoompu, K., Danese, A., Interlandi, M.,
2864-x. Mueller, M.F., Strobl, D.C., Zappia, L., Dugas, M., Colomé-Tatché, M.,
et al. (2022). Benchmarking atlas-level data integration in single-cell geno-
46. Hennig, B.P., Velten, L., Racke, I., Tu, C.S., Thoms, M., Rybin, V., Besir, H.,
mics. Nat. Methods 19, 41–50. https://doi.org/10.1038/s41592-021-
Remans, K., and Steinmetz, L.M. (2018). Large-scale low-cost NGS library
01336-8.
preparation using a robust Tn5 purification and tagmentation protocol. G3
(Bethesda) 8, 79–89. https://doi.org/10.1534/g3.117.300257. 56. Bauer, M., Vaxevanis, C., Bethmann, D., Massa, C., Pazaitis, N.,
47. Gong, H., Holcomb, I., Ooi, A., Wang, X., Majonis, D., Unger, M.A., and Wickenhauser, C., and Seliger, B. (2020). Multiplex immunohistochemistry
Ramakrishnan, R. (2016). Simple method to prepare oligonucleotide-con- as a novel tool for the topographic assessment of the bone marrow stem
jugated antibodies and its application in multiplex protein detection in sin- cell niche. Methods Enzymol. 635, 67–79. https://doi.org/10.1016/bs.mie.
gle cells. Bioconjug. Chem. 27, 217–225. https://doi.org/10.1021/acs.bio- 2019.05.055.
conjchem.5b00613. 57. Bingham, E., Chen, J.P., Jankowiak, M., Obermeyer, F., Pradhan, N.,
48. Paczulla, A.M., Rothfelder, K., Raffel, S., Konantz, M., Steinbacher, J., Karaletsos, T., Singh, R., Szerlip, P., Horsfall, P., and Goodman, N.D.
Wang, H., Tandler, C., Mbarga, M., Schaefer, T., Falcone, M., et al. (2018). Pyro: deep universal probabilistic programming. https://doi.org/
(2019). Absence of NKG2D ligands defines leukaemia stem cells and me- 10.48550/arXiv.1810.09538.
diates their immune evasion. Nature 572, 254–259. https://doi.org/10. 58. Finak, G., McDavid, A., Yajima, M., Deng, J., Gersuk, V., Shalek, A.K.,
1038/s41586-019-1410-1. Slichter, C.K., Miller, H.W., McElrath, M.J., Prlic, M., et al. (2015). MAST:
49. Wolock, S.L., Lopez, R., and Klein, A.M. (2019). Scrublet: computational a flexible statistical framework for assessing transcriptional changes and
identification of cell doublets in single-cell transcriptomic data. Cell Syst. characterizing heterogeneity in single-cell RNA sequencing data.
8, 281–291.e9. https://doi.org/10.1016/j.cels.2018.11.005. Genome Biol. 16, 278. https://doi.org/10.1186/s13059-015-0844-5.

Cell Stem Cell 30, 706–721, May 4, 2023 721

 ll
OPEN ACCESS Resource

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Antibodies
For a complete list of the antibodies used N/A N/A
in this study, see Table S1.
Biological samples
For a list of the biological samples used N/A N/A
in this study, see Table S3.
Chemicals, peptides, and recombinant proteins
DBCO-PEG5-NHS Ester Jena Bioscience CLK-CSTM
Cell Staining buffer Biolegend 420201
Human TrueStain FcX Biolegend 422302
TrueStain Monocyte Blocker Biolegend 426102
UltraPure BSA Thermo Fisher AM2616
DRAQ7 Biolegend 424001
Incucyte Caspase3/7 Red VWR International MSPP-4704
Homemade Tn546 CRG Protein N/A
Technologies Unit
N,N Dimethylformamide Sigma Aldrich D4551-250ML
Betain Lösung Sigma Aldrich B0300-5VL
Recombinant RNase Inhibitor TaKaRa 2313B
Maxima H Minus Reverse Transcriptase ThermoFisher EP0752
KAPA HiFi HotStart ReadyMix Roche KK2602
SCF Peprotech 300-07
Flt3-L Peprotech 300-19
TPO Peprotech 300-18
IL-3 Peprotech 200-03
IL-6 Peprotech 200-06
UM729 Stem Cell Technologies 72332
Critical commercial assays
Chromium Next GEM Single Cell 3’ GEM, 10x genomics PN-1000121
Library & Gel Bead Kit v3.1
Chromium Next GEM Single Cell 3ʹ Kit v3.1 10x genomics PN-1000268
Chromium Next GEM Chip G Single Cell Kit 10x genomics PN-1000120
Single Index Kit T Set A 10x genomics PN-1000213
Dual Index Kit TT Set A 10x genomics PN-1000215
CleanPCR beads CleanNA CPCR-0050
AmpureXP Beads Beckman Coulter A63881
SPRIselect beads Beckman Coulter B23318
Qubit High Sensitivity dsDNA Assay ThermoFisher Q32851
Agilent Bioanalyzer High Sensitivity Agilent 5067-4626
Opal seven-color IHC kit Akoya Biosciences NEL811001KT
Deposited data
Count tables and metadata Figshare Figshare: https://doi.org/10.6084/m9.figshare.20291628
Full input and output of CloneTracer model Figshare Figshare: https://doi.org/10.6084/m9.figshare.21982496
Additional data: Patient A.9 follow up (Figure S5) Figshare Figshare: https://doi.org/10.6084/m9.figshare.21982490
Additional data: Patient A.2 follow up (Figure S5) Figshare Figshare: https://doi.org/10.6084/m9.figshare.21982454
(Continued on next page)

e1 Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023

 ll
Resource OPEN ACCESS

Continued
REAGENT or RESOURCE SOURCE IDENTIFIER
Additional data: MutaSeq data (Figures 4D and 4E) Figshare Figshare: https://doi.org/10.6084/m9.figshare.21982424
All single cell RNA-seq datasets, raw data EGA EGA: EGAS00001007078
Reference data4 Figshare Figshare: https://doi.org/10.6084/m9.
figshare.13397987.v3
Reference data31 Zenodo Zenodo: https://doi.org/10.5281/zenodo.6496279
Experimental models: Organisms/strains
NOD.Prkdcscid.Il2rgnull (NSG) mice Jackson Laboratory 005557
Oligonucleotides
For a complete list of oligonucleotides N/A N/A
used in this study, see Table S4.
FISH probes: 6q21/8q24 MetaSystems D-5802-100-OG
FISH probes: 7cen/7q22/7q36 MetaSystems D-5043-100-TC
Software and algorithms
CloneTracer and Primer Design code Zenodo Zenodo:
ComplexHeatmap CRAN v. 2.6.2
FlowJo FlowJo, LLC v. 10.8.1, 10.6.1
ggplot2 CRAN v. 3.3.5
Htseq (https://pypi.org/project/HTSeq/) PyPI v. 2.02
mitoClone Zenodo Zenodo: https://doi.org/10.5281/zenodo.4443074
Mgatk (https://github.com/caleblareau/mgatk) Github v. 0.1.1
Pheatmap CRAN v. 1.0.12
PhISCS (https://github.com/sfu-compbio/PhISCS) Github v. 1.0.0
R CRAN v. 4.0.2
Seurat CRAN v. 4.3.0
Scrublet (https://github.com/swolock/scrublet) Github v. 0.2.3
Scmap Bioconductor https://doi.org/10.18129/B9.bioc.scmap
Scanorama (https://github.com/brianhie/scanorama) Github v. 1.7.3
Spectre (https://github.com/ImmuneDynamics/Spectre) Github v. 1.0.0
STAR (https://github.com/alexdobin/STAR) Github v. 2.5.4
Other
StemSpan SFEM media Stem Cell Technologies 09650

RESOURCE AVAILABILITY

Lead contact
Requests for further information, resources and reagents should be directed to and will be fulfilled by the lead contact, Lars Velten
([email protected]).

Materials availability
This study did not generate new unique reagents.

Data and code availability

d Datasets including processed and integrated gene expression data, cell type annotation, clonal assignments, metadata and
dimensionality reduction are publicly available as Seurat v3 objects through figshare. The DOI is listed in the key resources ta-
ble. To protect patient privacy and as requested by the relevant ethics boards, raw sequencing data is available from the Eu-
ropean Genome-Phenome Archive upon submitting a data access agreement. To obtain these data, contact the lead contact.
All accession numbers of data analyzed in this manuscript are listed in the key resources table.
d The implementation of the model and code for primer design and data processing of Optimized 10x libraries is deposited at
Zenodo and publicly accessible. The DOI is listed in the key resources table. Code can also be found at https://github.com/
veltenlab/CloneTracer.
d Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023 e2

 ll
OPEN ACCESS Resource

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Human subjects
Bone marrow samples from AML patients were obtained at the Heidelberg University Hospital after informed written consent using
ethic application number S-169/2017. For demographic characteristics of sample donors, see Table S3. All experiments involving
human samples were approved by the ethics committee of the University Hospital Heidelberg and were in accordance with the
Declaration of Helsinki.

Animals
NOD.Prkdcscid.Il2rgnull (NSG) mice were bred and housed under specific pathogen-free conditions in the central animal facility of
the German Cancer Research Center (DKFZ). Animal experiments were approved and performed in accordance with all regulatory
€sidium Karlsruhe). Immune compromised, healthy, female NSG mice 8-12 weeks
guidelines of the official committee (Regierungspra
of age and an average weight of 18-25 g were sublethally irradiated (175 cGy) 24 h before xenotransplantation assays.

METHOD DETAILS

Collection of bone marrow

Bone marrow aspirates were collected from iliac crest. Mononuclear cells were isolated by Ficoll (GE Healthcare, Chicago, Illinois,
USA) density gradient centrifugation and stored in liquid nitrogen until further use.

Panel, exome, and bulk ATAC sequencing

For bone marrow samples form cohort A, CD3- and CD3+ cells were sorted by FACS and subjected to exome sequencing as
described before.21 GATK best practices were followed. Mutect2 with Tumor with match normal option was used for the identifica-
tion nuclear mutations specific for each patient. We considered CD3- cells as tumor and CD3+ as normal. Results are summarized in
Table S3.
Additionally, samples A.1, A.3, A.5, A.6, A.7, A.11, A.12, A.13, A.15 were subjected to bulk ATAC sequencing to identify mitochon-
drial mutations. Again, Mutect2 with Tumor with match normal option was used to identify variants in the mitochondrial genome.
Results are summarized in Table S5.
Bone marrow samples from cohort B were sequenced at diagnosis time point with the Illumina TruSight Myeloid Sequencing Panel
(Illumina, San Diego, USA) to determine the mutation status of leukemia driver mutations.

Antibody-oligo conjugation
For markers where no commercial conjugates were available, azide-modified oligonucleotides were conjugated to purified anti-
bodies (anti-human CD166, Clone 3A6 (Biolegend, 343902); anti-human GPR56, Clone 4C3 (Biolegend, 391902)) by the use of a
DBCO-PEG5-NHS Ester (Santa Cruz Biotechnology, Dallas, USA) in a copper-free click reaction.47
In brief, azide-containing storage buffer of purified antibodies was exchanged to PBS (pH 8.5) using the Amicon Ultra-0.5 NMWL
30 kDa Centrifugal Filter (EMD Millipore, Billerica, USA).
100 mg of PBS-buffered antibody was incubated with 2mM DBCO-PEG5-NHS in a final reaction volume of 100mL for 30 minutes at
room temperature. The reaction was stopped by the addition of 100mM Tris HCl (pH 8) for 5 minutes at room temperature and non-
reactive DBCO-PEG5-NHS was removed using the Amicon Ultra-0.5 NMWL 30 kDa Centrifugal Filter.
Azide-modified oligonucleotides were reconstituted in PBS before adding 30pmol per 1mg DBCO-functionalized antibody. The
click reaction was conducted at 4 C for 18 hours. Unreacted oligonucleotides were removed using the Amicon Ultra-0.5 NMWL
50 kDa Centrifugal Filter and the final volume was adjusted to 100mL using PBS (pH 8.5).
Conjugation products were confirmed on Ethidiumbromide (EtBr) stained 2% agarose gels, Coomassie brilliant blue (CBB) stained
4-12% polyacrylamide gels and by absorbance spectroscopy.
Azide-modified oligonucleotides were purchased from Biolegio (Biolegio, Nijmegen, Netherlands) and contained an antibody-spe-
cific barcode (bold), a PCR handle (italic) and a capture sequence (underlined). * indicates a phosphorothioated bond to prevent
nuclease degradation:
CD166: 5’/Azide/CCTTGGCACCCGAGAATTCCACATTAACAGCGCCAACAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA*A*A
GPR56: 5’/Azide/CCTTGGCACCCGAGAATTCCATCATATCCGTTGTCCCAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA*A*A

CITEseq surface labeling and FACS sorting

Human bone marrow samples were thawed and stained using the CITEseq antibody pool (Table S1), as well as sorting antibodies
according to the BioLegend protocol https://www.biolegend.com/en-us/protocols/totalseq-a-antibodies-and-cell-hashing-with-
10x-single-cell-3-reagent-kit-v3-3-1-protocol
In cohort A, sorting was done using fluorophore-tagged antibodies from BioLegend (San Diego, USA) against human CD3 (clone
UCHT; 1:30), CD34 (clone 581; 1:100), and GPR56 (clone CG4; 1:20). FACS sorting of live bone marrow cells was performed on a BD
FACSAria equipped with a 70 mm nozzle to enrich for the following populations: CD3+, CD3-CD34+, CD3-CD34-, while aiming for a
representation of 25%, 50%, 25%. When insufficient CD34+ were available, a maximum of CD34+ cells were sorted and GPR56 was

e3 Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023

 ll
Resource OPEN ACCESS

used as a second sorting marker to enrich stem cells from the CD34- fraction. Population frequencies were recorded and accounted
for in quantitative analysis of the single-cell RNA-seq data set. Sorted cells were loaded onto the Next GEM Chip G for a targeted cell
recovery of 5000 cells following the manufacturer’s instructions (10x Genomics, CG000206 Rev D).
In cohort B, fluorophore-tagged antibodies against human CD34 (clone 581) and CD45 (clone HI30) (patients B.1, B.2, B.3), or
CD56 (clone QA17A16) and CD45 (patient B.4) were used to enrich for following populations:

B.1 B.2 B.3 B.4

d0 d15 d0 d15 d0 d15 d0 d21 d105
total BM/CD45+ 85% 99.6% 100% 100% 92% 100% 100% 100% 98%
CD45dim/CD34+ 15% 0.4% 0% 0% 8% 0% - - -
CD45dim/CD56+ - - - - - - 0% 0% 2%

Population frequencies were recorded and accounted for in quantitative analysis of the single-cell RNA-seq data set. In particular,
in analyses that investigate the absolute frequencies of cell types in bone marrow (Figure 5D) the frequency of the cell type was
computed per sorted population, and multiplied with the frequency of the sorting gates in the bone marrow sample.
In cases where different biological samples were combined in the same GEM generation run, cells were labeled additionally with
oligonucleotide coupled cell hashing antibodies (Biolegend, San Diego, USA). FACS sorting of live bone marrow cells was performed
using DRAQ7 (1:1000; Biolegend, San Diego, USA) and Incucyte Caspase3/7 Red (1:5000; VWR International, Radnor, Pennsylvania,
USA) on a BD FACSAriaTM Fusion equipped with a 100mm nozzle. Sorted cells were loaded onto the Next GEM Chip G for a targeted
cell recovery of 8000 cells following the manufacturer’s instruction (10x Genomics, CG000206 Rev D).

Single-cell RNA sequencing

cDNA libraries were generated using the 10x Genomics 3’ gene expression kit version 3.1 according to the manufacturer’s instruc-
tions (10x Genomics, CG000206 Rev D). At the cDNA amplification step (step 2.2 of the 10x Genomics protocol), additive primers for
amplification of the ADT and HTO libraries were added according to the manufacturer’s instructions (Biolegend protocol:
TotalSeqTM-A Antibodies and Cell hashing with 10x Single Cell 3’ Reagent Kit v3 or v3.1 (Single Index) Protocol, Step II).
Following cDNA amplification (10x Genomics protocol: step 2.3A), cDNA was split: 10 mL were used for generating Gene Expression
(GEX) libraries and 5mL were used for generating antibody-derived tags (ADT) and hashtag oligo (HTO) libraries, respectively, accord-
ing to manufacturer’s instructions (GEX: 10x Genomics protocol: CG000206 Rev D, Step 3; ADT and HTO: Biolegend protocol:
TotalSeqTM-A Antibodies and Cell hashing with 10X Single Cell 3’ Reagent Kit v3 or v3.1 (Single Index) Protocol, Step III). The remain-
ing material was used to construct mitochondrial and targeted mutation libraries.
Final GEX, ADT and HTO libraries were quantified by Qubit and QC’ed on the Bioanalyzer.
Final GEX and ADT libraries were sequenced on separate lanes on a NovaSeq (Cohort A) or HiSeq4000 (Cohort B) with a targeted
sequencing depth of 50,000-100,000 reads/cell (GEX) and 300 reads/antibody/cell (ADT), respectively. HTO libraries were
sequenced with a targeted sequencing depth of 4000 reads/cell on a NextSeq500.

Optimized 10x: Mitochondrial libraries

For the full-length amplification of mitochondrial cDNA, mitochondrial primers were pooled so that each mitochondrial primer is pre-
sent at a final concentration of 0.9mM (mito primer mix). See Table S4 for all primer sequences used in this protocol. 10 ng of amplified
cDNA was added to a PCR master mix containing 50 ml 2X KAPA HiFi HotStart ReadyMix (Roche), 4 ml 10 uM PartialRead1 primer,
2.5 ml mito primer mix, in a total volume of 100mL. PCR was run as follows: 1 cycle of 95C for 3 mins, 11 cycles of [98C for 20 secs, 67C
for 1 min, 72C for 1 min], and 1 cycle of 72C for 5 mins followed by a 4C hold. PCR product was then cleaned with 1.5X (v/v) CleanPCR
beads (CleanNA), followed by two washes of 80% ethanol, and eluted in 30 ml EB (Qiagen), after which it was quantified by Qubit and
QC’ed by running 1-2 ng of DNA on an Agilent Bioanalyzer High Sensitivity chip. Sample bioanalyzer traces after this step are shown
in Figure S1A.
Mitochondrial mutation libraries were then generated by tagmentation with an in-house produced wild-type transposase (Tn5).46
Briefly, transposome assembly and linker loading was carried out by adding 1 ml of 2 mg/ml Tn5 and 1 ml of annealed linker Tn5ME-B/
Tn5MErev to 9 ml of water followed by incubation at 23C whilst shaking at 300 RPM for 30 mins. Assembled transposome was then
diluted 1:100 with water. In our experience, it typically required four parallel tagmentation reactions to generate adequate yield for
sequencing for a given sample. In a single tagmentation reaction, 1.5 ng of cDNA were added to 10 ml of diluted Tn5 and 10 ml of
4X tagmentation buffer (40 mM Tris-HCl, pH 7.4; 40 mM MgCl2), 10mL DMF for a total of 40 mL. Tagmentation reaction in the PCR
was run as follows: 1 cycle of 55C for 3 mins, then a 10C hold. It is important that the PCR is already at 55C when the PCR tubes
are placed in the instrument. After tagmentation, 10 ml of 0.2% SDS was added to the tagmented mixture and incubated at room
temperature for 5 mins to neutralize the reaction. Once the transposase has been neutralized, the tagmented sample was added
to a PCR master mix of 54 ml 2X KAPA HiFi HotStart ReadyMix, 6 ml of 100% DMSO (Thermo Scientific), 10 ml of 10 mM Targeted
10X primer, and 10 ml of 10 mM N7XX primer. This reaction mix was split into two PCR tubes and PCR was run as follows: 1 cycle
of 72C for 3 mins, 1 cycle of 95C for 30 secs, 12 cycles of [98C for 20 secs, 60C for 15 secs, 72C for 30 secs], and 1 cycle of 72C

Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023 e4

 ll
OPEN ACCESS Resource

for 3 mins followed by a 10C hold. After the PCR, all reactions were pooled and underwent two rounds of successive bead cleanup. In
the first bead cleanup, 0.6X (v/v) CleanPCR beads were used, followed by two 80% ethanol washes and eluted in 50 ml EB. In the
second cleanup, 0.6X (v/v) CleanPCR beads were again used, followed by two 80% ethanol washes and eluted in 15 ml EB. The final
library was then quantified by Qubit and QC’ed on the Bioanalyzer. Representative bioanalyzer traces are shown in Figure S1B.

Optimized 10x: Targeted genotyping libraries

Nuclear mutations were selected from panel or exome sequencing data by choosing non-synonymous variants in expressed genes,
located < 1.5kb away from the end of the gene. Primers targeting mutations of interest were designed using a customized version of
the TAPseq Bioconductor package,18 https://github.com/veltenlab/CloneTracer/tree/master/primer_design
Four rounds of PCRs were then used to generate nuclear mutation libraries. The first three rounds of PCRs are gene-specific
nested PCRs and sequencing adaptors and indices were added in the last PCR.
In the first round of PCR (PCR1), 10 ng of amplified cDNA from 10x was added to a PCR master mix containing 2.5 ml of pooled outer
gene-specific primers (final concentration of each individual primer in the final pool 10 mM-100 mM), 5 ml of 1 mM Partial_Read1 primer,
20 ml of 5 M Betaine, 50 ml of 2X KAPA HiFi HotStart ReadyMix, and topped up to 100 ml with nuclease-free water. PCR with a heated
lid was run as follows: 1 cycle of 95C for 3 mins, 11 cycles of [98C for 20 secs, 67C for 1 min, 72C for 1 min], and 1 cycle of 72C for
5 mins followed by a 4C hold. PCR product was then cleaned with 1.5X (v/v) CleanPCR beads (CleanNA), followed by two washes of
80% ethanol, and eluted in 15 ml EB (Qiagen). After each round of post-PCR cleanups, PCR products were quantified by Qubit and
QC’ed by running 1-2 ng of DNA on an Agilent Bioanalyzer High Sensitivity chip. Example bioanalyzer traces are shown in Figure S1H.
In the second round of PCR (PCR2), 10 ng from PCR1 was added to a PCR master mix containing 2.5 ml of pooled middle gene-
specific primers, 5 ml of 1 mM Partial_Read1 primer, 20 ml of 5 M Betaine, 50 ml of 2X KAPA HiFi HotStart ReadyMix (Roche), and top-
ped up to 100 ml with nuclease-free water. PCR was run as follows: 1 cycle of 95C for 3 mins, 10 cycles of [98C for 20 secs, 67C for
1 min, 72C for 1 min], and 1 cycle of 72C for 5 mins followed by a 4C hold. Again, PCR product was cleaned with 1.5X (v/v) CleanPCR
beads (CleanNA), followed by two washes of 80% ethanol, and eluted in 30 ml EB (Qiagen). Example bioanalyzer traces are shown in
Figure S1I.
The third round of PCR (PCR3) was run separately for each target gene. 10 ng from PCR2 was added to a PCR master mix con-
taining 2.5 ml of pooled staggered gene-specific primers (concentration of each primer in the final pool: 25 mM), 5 ml of 1 mM
Partial_Read1 primer, 20 ml of 5 M Betaine, 50 ml of 2X KAPA HiFi HotStart ReadyMix (Roche), and topped up to 100 ml with
nuclease-free water. PCR was run as follows: 1 cycle of 95C for 3 mins, 10 cycles of [98C for 20 secs, 67C for 1 min, 72C for
1 min], and 1 cycle of 72C for 5 mins followed by a 4C hold. PCR product was cleaned with 1.5X (v/v) CleanPCR beads
(CleanNA), followed by two washes of 80% ethanol, and eluted in 30 ml EB (Qiagen).
Finally, libraries were uniquely indexed for each sample. To this end, 10 ng from PCR3 was added to a PCR master mix containing
2.5 ml of 10 mM SI primer, 2.5 ml of 10 mM RPI-N7XX primer (see Table S1), 50 ml of 2X KAPA HiFi HotStart ReadyMix (Roche), and
topped up to 100 ml with nuclease-free water. PCR was run as follows: 1 cycle of 95C for 3 mins, 8 cycles of [98C for 20 secs,
52C for 15 sec, 72C for 45 sec], and 1 cycle of 72C for 5 mins followed by a 4C hold. PCR product was cleaned with 1.5X (v/v)
CleanPCR beads (CleanNA), followed by two washes of 80% ethanol, and eluted in 15 ml EB (Qiagen). Example bioanalyzer traces
are shown in Figure S1J.

Plate-based single-cell RNA-seq (MutaSeq)

Defrosted bone marrow mononuclear cells were stained with following antibodies: Lineage antibodies (CD3, CD19, CD20, CD235a)
and additional antibodies (CD34, CD38, CD36, CD45RA, CD90, CD49f, NKG2DL48). The single cell index sort was performed on a BD
FACSAria Fusion (BD Biosciences) equipped with 355, 405, 488, 561, and 640 nm lasers and Lin- CD34+ cells were sorted into single
wells containing lysis buffer to enrich the stem cell compartment except for 3 rows per 384-well plate, in which CD34- cell populations
were sorted. After the sort plates were flash frozen and stored at -80 C until library preparation.
Primer design and single-cell RNA-seq was performed as described.21 See Table S4 for all primers used. Cells were sorted into
384-well instead of 96-well plates and therefore the reaction volumes were downscaled by a factor 2.5-5 depending on the reaction.
Lysis volume was 1.2 ml per well. For reverse transcription 2 ml of a buffer containing 0.1 ml Maxima H Minus Reverse Transcriptase
(200 U/ml), 0.6 ml 5x RT buffer (both Thermo Scientific), 0.07 ml Recombinant RNase Inhibitor (TaKaRa), 0.45 ml PEG 50%, 0.09 ml
100 mM Smart-seq2 TSO (Eurogentec) and 0.69 ml nuclease-free H2O were added and RT was performed for 90 min at 42  C followed
by enzyme inactivation at 70  C for 15 min. The PCR reaction was downscaled to 3 ml. The cDNA was cleaned up using a 0.9x volume
(5 ml) of AmpureXP Beads (Beckman Coulter) and tagmented using homemade Tn546 at a dilution of 1:50. cDNA was used at a con-
centration of 1-3 ng/ml and 0.4 ml was tagmented by addition of 1.2 ml of Tn5 dilution mixed 1:1 with 2x tagmentation buffer (20 mM
Tris-HCl pH7.5, 20 mM MgCl2, 50% DMF) at 55  C for 10 min and afterwards shifted to ice. 0.4 ml 0.1% SDS was added for inac-
tivation and incubated for 5 min on ice. PCR was performed by adding 2.7 ml of KAPA HiFi HS mastermix, 0.3 mL of DMSO and
0.4 mL each of the forward i5 and reverse i7 library primer at 3 mM. The PCR program was 72  C 3 min, 98 C 30 sec, 12 cycles of
[98 C, 20s ec, 63 C, 15 sec, 72 C 30 sec], and 72 C, 3 min. Libraries were pooled and purified with 0.9x AmpureXP Beads.

Genotyping of single cell derived cultures

Single cell cultures and genotyping were performed as described 21 with the following modifications: Bone Marrow mononuclear cells
from patient A.6 were stained with following antibodies (CD3, CD45RA, CD33, CD98, CD49f, CD38, CD11c, CD371 and HLA-DR).

e5 Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023

 ll
Resource OPEN ACCESS

Lin- or Lin-CD34 + single cells were index-sorted into U-bottom 96-well plates (Sarstedt) containing 100 ml StemSpan SFEM media
(Stem Cell Technologies). Media was supplemented with penicillin/streptomycin (100 ng/mL), UM729 (1 mM, Stem Cell Technologies)
and the following human cytokines (all from Peprotech): SCF (20 ng/mL), Flt3-L (20 ng/mL), TPO (50 ng/mL), IL-3 (20 ng/mL), IL-6
(20 ng/mL). After two weeks at 5% CO2 and 37  C, colonies were imaged by microscopy, and harvested in 12 ml buffer RLT (Qiagen)
for subsequent DNA isolation.

Raw 10x Genomics data processing

Gene expression data was processed using cellranger version 4.0.0 with default parameters for feature barcoding. Doublets were
removed using scrublet (v. 0.2.3).49 For cohort A, cells with <1800 genes detected or >10% mitochondrial reads were removed.
For cohort B, cells with <1000 genes detected or >40 % mitochondrial reads were removed. The data from the healthy reference
individual (C.3) was downloaded from https://doi.org/10.6084/m9.figshare.13397987.v3 and not subjected to further quality filters.
Mitochondrial libraries were processed following the DropSeq standard workflow50 except that reads were aligned to the mito-
chondrial genome (GRCh38). Consensus mitochondrial reads were called using the fgbio tool CallMolecularConsensusReads (v.
1.3.0). Only reads from cell barcodes which were detected in the gene expression dataset were used for the downstream analysis.
Nucleotide counts were extracted for each single cell using pysam (v. 0.15.3). The final output of the workflow is a list of single-cell
matrices in which for each position of the mitochondrial genome the number of A,T,C and Gs UMIs are stored. Mitochondrial variants
were identified as previously described,21 and using bulk ATAC sequencing, where available (most of cohort A). The workflow was
implemented in snakemake51 and can be found in https://github.com/veltenlab/CloneTracer/tree/master/library_processing/
mitochondria
Nuclear SNV libraries were processed similarly to the mitochondrial libraries with the difference that reads were aligned to the com-
plete human genome (GRCh38). Only reference and alternative alleles (identified by exome or panel sequencing) were considered for
the final count table. Due to the high number of PCR amplification steps, only UMIs supported by at least two reads were included in
the analysis. The final count table contains the number of reference and alternative UMIs for each single cell and targeted mutation.
The workflow to process nuclear SNVs libraries was written in snakemake and can be found in https://github.com/veltenlab/
CloneTracer/tree/master/library_processing/nuclear-snv

Analysis of single cell gene expression data

For projecting single cell data onto a reference atlas of healthy bone marrow, we used a workflow based on scmap52 as described.4
Sample code for reference atlas projection is available at https://git.embl.de/triana/nrn/-/tree/master/Projection_Vignette. Thereby,
we obtained uMAP coordinates, cell type labels, and myelocyte pseudotime, where applicable.
For unsupervised integration of all data sets, scanorama53 was used with default parameters to integrate across the three cohorts
A, B and C, using the cohort as the batch. Scanorama components were then imported into Seurat and uMAP, nearest neighbor
graphs, and clustering were computed using the default Seurat pipeline with default parameters.54 scanorama was selected based
on a systematic comparison study,55 where it was described as a method that maintains biologically true difference between sam-
ples, which we considered relevant in the context of a highly heterogeneous disease such as AML.
To illustrate this, in Figure S2D, cluster C15, C16 and C32 all resemble HSCs/MPPs when projected to a healthy reference.
Compared to other immature myeloid clusters all cells from these clusters express stem cell markers (CD34, MHC class II) and
lack markers of myeloid commitment (MPO, AZU1); but they also differ in the expression of genes that are usually co-expressed
in HSCs1 such as MECOM (exclusive to cluster C16), HOX genes (strongly overexpressed in cluster C32), or, in the case of cluster
C15, displayed a strong interferon response signature. Since these are real biological, and not technical differences, we think it is
important that they are represented in the uMAP and unsupervised clustering. Thereby data integration by scanorama is highly com-
plementary to the projection to a healthy reference, that we employ e.g. in the context of Figures 3A, 3G, and 3H: This two-tiered
approach (projection and weak batch integration) allows us to identify biologically different cell states that all resemble e.g.
MPPs, but vastly differ in the (aberrant) expression of key genes.
As a further validation of the integration strategy employed, we included a healthy control individual in each cohort. We then inte-
grated all three healthy individuals (A.0, B.0 and the Reference individual from4) using the same unsupervised data integration steps
used to generate the main figures of the manuscript, with identical parameters. The resulting uMAP (Figure S2E) demonstrates that
our data integration strategy effectively accounts for technical differences between cohorts.

Dormancy score calculation

To compute a dormancy score for each individual cell, we first selected genes differentially expressed in Zhang et al.32 (adjusted
p-value <0.01) and the dHSC gene list from Cabezas-Wallscheid et al.33 (adjusted p-value <0.05). We then normalized the gene
expression count matrix by library size and centered and scaled the data. We computed principal components for cluster C6 cells
from all patients with <100 of cells in this population using the prcomp() function in R. The dormancy score corresponded to the first
principal component. To compute the score in other cells we used the function predict() with the principal component loadings
computed as described above as a model and the scaled gene expression data as new input.

Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023 e6

 ll
OPEN ACCESS Resource

Analysis of DNAseq from single-cell derived colonies

Raw sequencing reads were aligned to the human genome using STAR (v. 2.5.4). Nucleotide count tables were generated from sin-
gle-colony BAM files using the function baseCountsFromBamList from the package mitoClone (v.1.0). For each mutation, colonies
were labelled as mutant when > 5% of the reads were mutant and healthy otherwise. If the mutation was not covered it was labelled
as dropout. The binarized table (colonies x mutations) was the input to PhISCS41 which was ran with default parameters to infer the
clonal hierarchy.

Processing of MutaSeq scRNAseq data

Raw gene expression data was aligned using STAR (v. 2.5.4) and count matrices were generated using htseq (v. 2.02) with default
parameters. Only cells with > 2000 genes detected and <10% mitochondrial reads were kept for downstream analysis. Gene expres-
sion data for all patients was integrated using scanorama.53 The 5000 most variable genes were included. Genes which were ampli-
fied with mutation-specific primers were also excluded from the integration process. Each patient was used as batch key. The first
100 scanorama components were used to compute the integrated uMAP following the default Seurat pipeline. Reference and mutant
counts for SNVs and mtSNVs were obtained using the function baseCountFromBamList from the mitoClone package (v.1.0)21

Raw data processing of MAESTER data

Raw fastq files from the MAESTER library of a human clonal hematopoiesis sample were downloaded from SRA (SRR15598777) and
processed as described.14 In brief, 24bp primer sequences were trimmed from the 5’ of read2 fastq file using homertools (v. 4.11).
Read 2 was tagged with cell and umi barcodes and aligned to the complete human genome using STAR (v.2.5.4). Only reads from
cells present in the final Seurat object (downloaded from https://www.dropbox.com/s/vna1k3k7khazd7j/BPDCN712_Seurat_Final.
rds?dl=0) were kept for downstream analysis. mgatk (v. 0.1.1) was used to obtain single-cell nucleotide count matrices with default
parameters and –mr = 3.

Fluorescent In Situ Hybridization

Human bone marrow samples were thawed and stained with fluorophore-tagged antibodies against CD45, CD3, CD49f, CD11c,
CD14 and CD34 as described above (CITEseq surface labeling, FACS sorting and GEM generation). For antibody clones and titrated
amounts, see Table S1. Cells were collected by FACS sorting on a BD FACSAriaTM,or BD FACSAriaTM Fusion, respectively, each
equipped with a 100mm nozzle. Sorted cells were fixed on glass slides in methanol/acetic acid. Hybridization was performed accord-
ing to the manufacturer’s instructions by using FISH probes for chromosome regions 6q21/8q24 and 7cen/7q22/7q36
(MetaSystems, Altlussheim, Germany), respectively. Interphase nuclei were validated using an automated scanning system (Applied
Spectral Imaging, Edingen/Neckarhausen, Germany).

Tissue microarrays
The frequency of different cell subsets in the bone marrow microenvironment in AML patients was analyzed by multispectral imaging
(MSI). Formalin-fixed and paraffin embedded (FFPE), decalcified bone marrow samples were stained as described elsewhere.56 The
marker panel used for staining included antibodies directed against CD34, CD14, CD11c, CD49f. For antibody clones and dilutions,
see Table S1. All primary antibodies were incubated for 30 min. Tyramide signal amplification (TSA) visualization was performed using
the Opal seven-color IHC kit containing fluorophores Opal 520, Opal 540, Opal 570, Opal 690 (Akoya Biosciences., Marlborough, MA,
USA), and DAPI. Stained slides were imaged employing the PerkinElmer Vectra Polaris platform. To unify the spatial distribution
analysis, 3 20 MSI fields (1872 3 1404 pixel, 0.5 mm/pixel) were analyzed. Cell segmentation and phenotyping of the cell subpop-
ulations were performed using the inForm software (PerkinElmer Inc., USA). The frequency of all immune cell populations analyzed
and the cartographic coordinates of each stained cell type were obtained.

Large cohort flow cytometry analysis

Human BM samples were stained as described above and analyzed using the BD Symphony (see Table S1 for a list of antibodies
used). Cells were pre-analyzed using FlowJo v10.8.1. Doublets and dead cells were excluded, as well as artefacts using a time
gate. Remaining cells were exported using the channel values and imported into R. Further, the package Spectre v1.0.0 was
used for batch correction, clustering, dimension reduction and visualization following the ‘‘discovery workflow with batch alignment
using CytoNorm’’. Using the summary table of the package, CD11c expression was investigated on CD34+ and CD14+ cell clusters.

Xenotransplantations
Female NSG mice 8-12 weeks of age were sublethally irradiated (175 cGy) 24 h before xenotransplantation assays. FACS sorted
primary AML samples were injected into the femoral BM cavity of sublethally irradiated mice. Mice were daily monitored, and femur
bone marrow aspirates were taken at 16 weeks to determine engraftment and lineage potential. Human leukemic engraftment in
mouse BM was evaluated by flow cytometry using anti-human-CD45-FITC (clone HI30), anti-human-CD34-BUV395 (clone 581),
anti-human-CD19-FITC (clone HIB19), anti-human-CD33-PE-Cy7 (clone WM53), CD3-BV510 (clone OKT3), and anti-mouse-
CD45-Alexa700 (clone 30-F11).

e7 Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023

 ll
Resource OPEN ACCESS

QUANTIFICATION AND STATISTICAL ANALYSIS

CloneTracer model
See Methods S1 for a full description of the CloneTracer model. Posterior predictive checks57 were used to determine if the data met
the assumptions of the statistical model, as detailed in the Methods S1.

Differential expression testing

For differential expression testing of surface antigens, we used Wilcoxon tests following library size normalization. For differential
expression testing of RNA, we used MAST.58 In all cases, comparisons were performed separately by patients, and the number
of patients where the change was significant was used as an overall measure of significance and consistency.

Data visualization
All plots were generated using the ggplot2 (v. 3.3.5), ComplexHeatmap (v. 2.6.2) and pheatmap (v. 1.0.12) packages in R 4.0.2 or
FlowJo (v. 10.6.1, BD). Boxplots are defined as follows: the middle line corresponds to the median; the lower and upper hinges corre-
spond to first and third quartiles, respectively; the upper whisker extends from the hinge to the largest value no further than 1.5X the
inter-quartile range (or the distance between the first and third quartiles) from the hinge and the lower whisker extends from the hinge
to the smallest value at most 1.5X the inter-quartile range of the hinge. Data beyond the end of the whiskers are called ‘outlying’ points
and are plotted individually.
Detail on statistical tests used in the different figures and definition of relevant summary statistics are included in the figure legends.

ADDITIONAL RESOURCES

Interactive 2D and 3D versions of most uMAPs from this paper are available at https://veltenlab.crg.eu/clonetracer/.

Cell Stem Cell 30, 706–721.e1–e8, May 4, 2023 e8