Single-Cell Mass Cytometry of Differential Immune and Drug Responses Across a Human Hematopoietic Continuum

Baxter Laboratory in Stem Cell Biology, Department of Microbiology and Immunology, Stanford University, Stanford, CA 94305, USA.
Science (Impact Factor: 33.61). 05/2011; 332(6030):687-96. DOI: 10.1126/science.1198704
Source: PubMed
ABSTRACT
Flow cytometry is an essential tool for dissecting the functional complexity of hematopoiesis. We used single-cell “mass cytometry”
to examine healthy human bone marrow, measuring 34 parameters simultaneously in single cells (binding of 31 antibodies, viability,
DNA content, and relative cell size). The signaling behavior of cell subsets spanning a defined hematopoietic hierarchy was
monitored with 18 simultaneous markers of functional signaling states perturbed by a set of ex vivo stimuli and inhibitors.
The data set allowed for an algorithmically driven assembly of related cell types defined by surface antigen expression, providing
a superimposable map of cell signaling responses in combination with drug inhibition. Visualized in this manner, the analysis
revealed previously unappreciated instances of both precise signaling responses that were bounded within conventionally defined
cell subsets and more continuous phosphorylation responses that crossed cell population boundaries in unexpected manners yet
tracked closely with cellular phenotype. Collectively, such single-cell analyses provide system-wide views of immune signaling
in healthy human hematopoiesis, against which drug action and disease can be compared for mechanistic studies and pharmacologic
intervention.

Full-text

Available from: Peng Qiu
Single-Cell Mass Cytometry of Differential
Immune and Drug Responses Across
a Human Hematopoietic Continuum
Sean C. Bendall,
1
* Erin F. Simonds,
1
* Peng Qiu,
2
El-ad D. Amir,
3
Peter O. Krutzik,
1
Rachel Finck,
1
Robert V. Bruggner,
1,7
Rachel Melamed,
3
Angelica Trejo,
1
Olga I. Ornatsky,
4,5
Robert S. Balderas,
6
Sylvia K. Plevritis,
2
Karen Sachs,
1
Dana Peer,
3
Scott D. Tanner,
4,5
Garry P. Nolan
1
Flow cytometry is an essential tool for dissecting the functional complexity of hematopoiesis. We used
single-cell mass cytometry to examine healthy human bone marrow, measuring 34 parameters
simultaneously in single cells (binding of 31 antibodies, viability, DNA content, and relative cell size). The
signaling behavior of cell subsets spanning a defined hematopoietic hierarchy was monitored with 18
simultaneous markers of functional signaling states perturbed by a set of ex vivo stimuli and inhibitors.
The data set allowed for an algorithmically driven assembly of related cell types defined by surface
antigen expression, providing a superimposable map of cell signaling responses in combination with drug
inhibition. Visualized in this manner, the analysis revealed previously unappreciated instances of both
precise signaling responses that were bounded within conventionally defined cell subsets and
more continuous phosphorylation responses that crossed cell population boundaries in unexpected
manners yet tracked closely with cellular phenotype. Collectively, such single-cell analyses provide
system-wide views of immune signaling in healthy human hematopoiesis, against which drug action
anddiseasecanbecomparedformechanisticstudies and pharmacologic intervention.
F
luorescence-based flow cytometry has
been fundamental to the discovery and
definition of major and minor cell subsets
of the immune system. Although the outline of
hematopoiesis is generally understood (1), a com-
prehensive framewo rk of its system-wide proper -
ties remains to be determined (2). Technological
developments in flow cytometry and cell sorting
[the introduction of new fluorophores, such as
quantum dots (3)] have paralleled appreciation
of the compartmentalization of function in the
hematopoietic system and contributed to diverse
fields, including immunology, stem cells (4, 5),
HIV (6), cancer (7), transcription (8, 9), intra-
cellular signaling (10, 11), apoptosis, cell cycle
(12), and development of cytometry-based clin-
ical diagnostics (13, 14 ). However, use of flow
cytometry remains practically confined to the
measurement of 6 to 10 simultaneous param-
eters (15). Analysis at the 11- to 15-parameter
range is possible but limited by compensation
needed to correct for spectral overlap that can
create a source of confounding variability (16 ).
We used transition element isotopes not nor-
mally found in biological systems as chelated anti-
body tags in atomic mass spectrometric analysis of
single cells to create a detailed response profile
of the healthy primary human hematopoietic sys-
tem with 34 simultaneously measured cellular
parameters . This allowed us to take full advantage
of the measurement resolution of mass spectrom-
etr y and ap p ly it t o singl e -c el l analy si s. Because
the method is largely unhampered by interference
from spectral overlap, it allows for the detection
of considerably more simultaneous parameters
than does traditional flow cytometry (17, 18).
Combined with its quantitative nature, atomic mass
spectrometry measurement creates a platform with
which to conduct multiplexed measurement of
single-cell biological parameters that can exhibit
vastly different dynamic ranges during signaling
or over time (such as signaling changes indicated
by shifts in protein phosphorylation).
We simultaneously measured 34 parameters
in each single cell in human bone marrow (BM)
samples to provide an in-depth analysis of normal
human hematopoietic and immunological signal-
ingoverlaidontoadetailed template of cell phe-
notype. Cell subsetspecific signaling phenotypes
of drug action in the face of clinically meaningful
physiologic stimuli were localized to pathway and
cell-specific boundaries, with examples in B cell
signaling shown. These provide a system-wide
view of signaling behaviors, expanding our view
of drug action while allowing us to limit the func-
tions that certain drugs might have on complex
tissues. Given that this technology can reason-
ably be expected to allow for as many as 100 pa-
rameters per cell (18, 19), it affords an opportunity
to increase our understanding of cell typespecific
signaling responses in complex, distributed or-
gans such as the immune system.
Performance assessmentofmasscytometry.
The workflow for mass cytometry is comparable
with that of fluorescence flow cytometry (Fig. 1A).
Antibodies coupled to distinct, stable, transition
element isotopes were used to bind target epitopes
on and within cells. Cells, with bound antibody-
isotope conjugates, were sprayed as single-cell
droplets into an inductively coupled argon plasma
(created by passing argon gas through an induc-
tion coil with a high radio-frequenc y elec tric cur-
rent) at approximately 5500 K. This vaporizes
each cell and induces ionization of its atomic con-
stituents. The resulting elemental ions were then
sampled by a time-of-flight (TOF) mass spectrom-
eter and quantified. The signal for each transi-
tion element isotope reporter was integrated as
each cells constituent ions reached the detector.
Currently, TOF sampling resolution enables the
measurement of up to 1000 cells per second.
We compared mass cytometry with conve ntional
nin e- p arameter fluorescence flow cytometry in
analysis of cytokine signaling through responses
in human peripheral blood mononuclear cells
(PBMCs) from two healthy donors (Fig. 1, B to E,
and fig. S1). Seven surface antigens (CD3, CD4,
CD8, CD45RA, CD56, CD20, and CD33) and
two intracellular phosphoprotein epitopes [phos-
phorylated signal transducer and activator of
transcription 3 and 5 (pST AT3 and pST AT5)] were
measured by means of fluorescence cytometr y on
two human PBMC samples treated with interleukin-2
(IL-2), IL-6, IL-10, granulocyte-monocyte colony
stimulating factor (GM-CSF), or interferon-a
(IFNa) to measure cytokine-mediated signaling
responses in specific cell subsets. In traditional
flow cytometry, forward scatter (FSC) and side
scatter (SSC) measurements of laser light are used
to detect the presence of a cell and to trigger the
electronics in order to collate information as a
cell event (the window of time during which a
cell is measured). Because FSC and SSC are not
currently implemented on the CyTOF platform,
alternative parameters providing analogous utility
were included to assist with the discrimination of
single-cell events: (i) an antibody to the surface
epitope CD45 (expressed on most cells measured
in this study), (ii) a metal-encoded DNA inter-
calator to identify nucleated cells (20), and (iii) a
derived parameter (cell length) indicating the
duration of each cells measurement window (18).
Fluorescence (Fig. 1B and fig. S1A) and mass
(Fig. 1C and fig. S1B) cytometry analysis pro-
vided comparable results when analyzed via tra-
ditional dot plots (fig. S2). Pertinent qualities,
such as reduced CD45RA expression on CD4
+
T cells relative to that on CD8
+
cells, were re-
produced between platforms (Fig. 1, B and C).
Despite use of different metrics for identifying
cell events, both platforms yielded quantitatively
similar frequencies (P < 0.000 001) for 12 man-
ually gated cell populations in the parallel anal-
ysis of two separate door samples (fig. S1C and
table S1). Patterns of specific induction of STAT
protein phosphorylation within the CD4
+
CD45RA
+
RESEARCH ARTICLE
1
Baxter Laboratory in Stem Cell Biology, Department of Micro-
biology and Immunology, Stanford University, Stanford, CA 94305,
USA.
2
Department of Radiology, Stanford University, Stanford, CA
94305, USA.
3
Department of Biological Sciences, Columbia Uni-
versity, New York, NY 10027, USA.
4
University of Toronto, Toronto,
ON M5S 3H6, Canada.
5
DVS Sciences, Markham, ON L3R 6E7,
Canada.
6
BD Biosciences, San Diego, CA 95131, USA.
7
Biomedical
Informatics Program, Stanford University, Stanford, CA 94305, USA.
*These authors contributed equally to this work.
To whom correspondence can be addressed. E-mail:
gnolan@stanford.edu
www.sciencemag.org SCIENCE VOL 332 6 MAY 2011 687
on May 14, 2011www.sciencemag.orgDownloaded from
Page 1
pSTAT3 pSTAT5
123456789 123456789
Population:
BC
E
-5.1 0 1.5
Fluorescence
Cytometry
Mass
Cytometry
pSTAT5
Unstimulated
p
STAT3
IL-2 IL-10
D
Fluorescence
Cytometry
Mass
Cytometry
α
IFN
Unstim
IL-2
IL-6
IL-10
GM-CSF
α
IFN
Unstim
IL-2
IL-6
IL-10
GM-CSF
α
IFN
A
ICP-MS
Elemental Analysis
Nebulize Single-Cell Droplets
Integrate
Signal
Antibodies
Labeled with
Elemental Isotopes
Upload
.FCS
Files
Cytobank.org
Isotope A
Isotope B
2D Plots Expression &
Fold-Change
SPADE Analysis
Mass Cytometer
Mass
Cell 1
Cell 2
Cell 3
**
-10
2
10
2
10
3
10
4
10
5
0
-10 10 10
3
10
2
10
4
0
-10
10
10
3
10
2
10
4
0
-10
2
10
2
10
3
10
4
10
5
0
FSC
10
5
10
4
0
CD8
CD45RA
CD4
CD45RA
CD33
SSC
CD3
CD20
CD56
CD3
1.CD33
Myeloid
2.CD20
B Cells
7.NKT
Cells
NK
Cells
CD8
T Cells
CD4
T Cells
Fluorescence Cytometry
98
3
46
5
SSC
0
50K
100K
150K
200K
250K
0
50K
100K
150K
200K
250K
0 50K 100K 150K 200K 250K
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
-10
2
10
2
10
3
10
4
10
5
0
0
25
50
75 100 125
Cell Length
DNA
CD8
CD45RA
CD4
CD45RA
CD33 CD3 CD56
CD8
T Cells
CD4
T Cells
Mass Cytometry
CD45
CD20
CD3
1.CD33
Myeloid
2.CD20
B Cells
7.NKT
Cells
NK
Cells
98
3
46
5
-10 10 10
3
10
2
10
4
0
-10 10 10
3
10
2
10
4
0
-10 10 10
3
10
2
10
4
0
-10 10 10
3
10
2
10
4
0
-10 10 10
3
10
2
10
4
0
-10
10
10
2
10
3
10
4
0
-10
10
10
2
10
3
10
4
0
-10
10
10
2
10
3
10
4
0
-10
10
10
2
10
3
10
4
0
-10
10
10
2
10
3
10
4
0
-10
10
10
2
10
3
10
4
0
*
Fig. 1. Mass cytometry profiling of immune cell response patterns. (A)Work-
flow summary of mass cytometry analysis. Cells are stained with epitope-specific
antibodies conjugated to transition element isotope reporters, each with a dif-
ferent mass. Cells are nebulized into single-cell droplets, and an elemental mass
spectrum is acquired for each. The integrated elemental reporter signals for each
cell can then be analyzed by using traditional flow cytometry methods as well as
more advanced approaches such as heat maps of induced phosphorylation and
tree plots. (B and C) Representa tive antibody surface-staining results and cell
population definitions (gating)for(B)fluorescenceand(C)masscytometry
analysis of fixed PBMCs from the same donor. Replicate analysis of a second
donor is provided in (21) (Fig. S1A and S1B). *Pearson correlation between
frequencies measured by fluorescence or mass cytometry, including both donors
(r =0.99,P < 0.000001, two-tailed t test) (table S1 and fig. S1C). (D) Induction
of STAT3 and 5 phosphorylation by various ex vivo stimuli in naive CD4
+
CD45RA
+
T cells [(B) and (C), red boxes] as measured by (top) fluorescence and (bottom)
mass cytometry. Red arrows indicate the expected shift along the STAT
phosphorylation axes. (E) Heatmap summary of induced STAT phosphorylation
in immune populations from the PBMC donor defined in (B) and (C) [column
headers refer to blue polygons in (B) and (C)]. Responses to the indicated stimuli
in each row were measured by (top) fluorescence and (bottom) mass cytometry.
Color scale indicates the difference in log
2
mean intensity of the stimulated
condition compared with the unstimulated control. Signaling responses of a
second donor are provided in (21) (fig. S1D). **Pearson correlation between
signaling induction measured by fluorescence or mass cytometry, including
both donors [pSTAT3: r =0.92;P < 0.000001, two-tailed t test (fig. S1E);
pSTAT5: r = 0.89, P < 0.000001, two-tailed t test] (figs. S1E and S1F).
6 MAY 2011 VOL 332 SCIENCE www.sciencemag.org
688
RESEARCH ARTICLE
on May 14, 2011www.sciencemag.orgDownloaded from
Page 2
T cell population demonstrated that both plat-
forms could equivalently det ec t pSTAT 3, p STAT5,
and dual pSTAT3-pSTAT5 responses to IL-10,
IL-2, and IFNa, resp ec tiv e ly (Fig. 1D). One qual-
itative difference between the two platforms was
the mathematical correction requ i r e d to add r e ss
spectral overlap in the fluorescence data (termed
compensation), a procedure not required with
the atomic mass spectrometer. A second major
distinction is the absence of cell-dependent back-
ground signal in the mass cytometry data. Thus,
although laser-based flow cytometry detects sig-
nals from cellular autofluoresc e n ce , nona c t i v at e d
cel l s had mas s cyto m e t r ic phosphoprotein inten-
sities near zero, indicating very little background
antibody binding. This manifests in atomic mass
spectrometry as a narrow grouping of cell events
at the low end of the dot plot axes. Qualitatively
and quantitatively (P < 0.000001) similar patterns
were revealed by means of fluorescence-based flow
cytometry or mass cytometry in terms of magnitude
of the pST AT3 and pST AT5 responses in cell pop-
ulations across two healthy peripheral blood sam-
ples (Fig. 1E and fig. S1, D to F). An overview of
the antibody quality control with testing on cell
lines, human PBMCs, and bone marrow is shown
in fig. S11 . T aken together , mass cytometry and
traditional fluorescence based approaches can pro-
duce results with equivalent informational value.
Organization and analysis of high-dimensional
single-cell data. Taking advantage of the in-
creased dimensionality of mass cytometry, we
prepared a set of reagents to capture a system-
wide view of immune cell types from a replicate
analysis of bone marrow mononuclear cells from
two healthy human donors. Thirty-one distinct
transition element isotopes were used to label
two antibody-staining panels for the study of
healthy human bone marrow mononuclear cells.
[Data are publicly available at Cytobank (www.
cytobank.org/nolanlab). An immunophenotyping
panel was designed that monitored 13 core
surface markers and 18 subset-specific cell-surface
markers to allow identification of human hema-
tologic cell types. A functional panel contained
the 13 core surface markers and also 18 intra-
cellular epitopes that reflect intracellular signaling
states, such as phosphorylation status of kinase
substrates (21). These complementary panels al-
lowed simultaneous biochemical analysis of intra-
cellular signaling in rare and diverse cell subsets
that were identified through in silico merging
of the data. Intracellular signaling responses were
de termined by treating cells ex vivo with modu-
lators such as cytokines, small molecules, or com-
binations thereof. Perturbation analysis has proven
useful in causality determinations for signaling
at the single-cell level (11, 2225) and was ap-
plied here to enable cell subsetspecific response
profiles. An additional three parametersaDNA
intercalator , cell length, and a cell viability dye
(21)were included in the analysis panels, creating
a total of 34 parameters in each. With an overlap
of 13 core surface antibodies between the two anal-
ysis panels and the three shared additional cell
features, a combined total of 52 unique single-cell
parameters were measured. The resulting single-
cell data set of bone marrow cells captured a
snapshot of the cell types present and their cor-
responding regulatory signaling responses through-
out development from early human hematopoietic
progenitors to lineage-committed cells.
A central dogma of immunology is that cells
at different stages of maturation can be char-
acterized by the expression of unique sets of
proteins on the cell surface. Such cluster of dif-
ferentiation (CD) markers are routinely used
for flow cytometric identification of cell popula-
tions. Although it is convenient to think of cells
in different stages of development as having dis-
tinct, regimented profiles, hematopoiesis fre-
quently manifests as a continuum of CD marker
expression connecting the cellular lineage stages
(26). Although cells might pause at recognized
stages of development to which we ascribe cer-
tain phenotypes, cells also pass through transient
intermediate states that connect parent popula-
tions to their progeny . As they proceed from one
stage of development to the next, CD marker
sets rise and fall in accordance with programmed
differentiation and environmental contexts. A con-
ventional display of the relationships between
the 31 cell surface markers measured here on
human bone marrow would require greater than
450 biaxial dot plots (fig. S3), making a com-
prehensive interpretation of the underlying cel-
lular progression unwieldy, if not impossible.
We hypothesized that the inherent similarity
of cell stages and continuity of the transitions
between cell differentiation states could be used
to organize high-dimensional data into ordered,
continuous clusters of similar cell phenotypes
that, when projected on a two-dimensional (2D)
plane, would convey the relatedness of these cells
in a higher dimensional space. We leveraged pro-
gressive changes in CD marker expression to
organize bone marrow cells in an unsupervised
manner , creating a tree-like scaffold for visual-
ization of high-dimensional intracellular sig-
naling behaviors in various cell types present
during hematopoietic development in the bone
marrow (27, 28). To accomplish this, we used
SPADE (spanning-tree progression analysis of
density-normalized events), a density normalization,
agglomerative clustering, and minimum-spanning
tree algorithm to distill multidimensional single-
cell data down to interconnected clusters of rare,
transitional, and abundant cell populations, which
were organized and di sp laye d as a 2D tree pl o t
(Fig. 2A). Such a tree plot from healthy bone
marrow represented the clustered expression of
the cell-surface antigens t ha t were used to bui ld
the tree in 13-dimensional space on the basis of
the core surface markers conserved between our
two 34-parameter analysis panels (CD3, -4, -8,
-1 1b, -19, -20, -33, -34, -38, - 45, -45RA, -90,
and -123) (Fig. 2B). Each node of the plot en-
compasses a cluster of cells that were pheno-
typically similar in the 13-dimensional space
defined by the core surface markers. The ap-
proach uses a minimum-spanning tree algorithm,
in which each node of cells is connected to its
most related node of cells as a means to con-
vey the relationships between the cell clust e rs .
The number of nodes and ultimately their bound-
aries is driven by a user-definable value (21).
Each node describes an n-dimensional boundary
encompassing a population of phenotypically
similar cells. When connected via the minimum
spanning tree, this provides a convenient ap-
proach to map complex n-dimensional relation-
ships into a representative 2D structure.
As such, related nodes could be mapped into
traditionally described immunological cell pop-
ulations as determined by the localized expression
patterns of 13 directly measured surface markers
(Fig. 2, B to E, and fig. S4A). A summary of evi-
dence supporting these annotations and bounda-
ries can be found in table S2. For instance, T cell
populations were annotated on the far right branch
of the tree plot based on the high expression of
CD3 (Fig. 2C, bright red). The T cell markers CD4
and CD8 were expressed in mutually exclusive
clusters but overlapped with CD3 expression. Den-
sity normalization enabled the display of rare cell
types, such as CD34
+
progenitor cells, in the same
space as the more abundant differentiated cell types
(Fig. 2E). The unsupervised orga nization of pheno -
ty p i ca l l y related cell types into adjacent branches,
such as CD4 and CD8 T cells (Fig. 2C), mature and
im mature B cells (Fig. 2D), and different clusters
of myeloid cells (Fig. 2E) collectively illustrates
that the algorithmic ordering of surface marke r
similarity can objectively or ganize cell types into
physiologically relevant compartments.
Although they were not used in the tree-building
step, the 18 additional surface markers from the
immunophenotype-staining panel were used to
confirm and refine cell subset annotations (Fig.
2F and fig. S4B). These markers were overlaid in
an unsupervised fashion onto the existing tree by
assigning each cell from the immunophenotyp-
ing experiment to whichever node contained
analogous cells from the functional data set ac-
cording to the expression of the shared 13 core
surface markers in the registration space. The ac-
curacy of this automated overlaying approach is
supported by the agreement of multiple natural
killer (NK), monocyte, and B cell markers that
localized to the appropriate cell populations (Fig.
2F), even though they were not used to direct the
trees original organization. Although the tree
structure derived from bone marrow data reca-
pitulates many features of hematopoietic organi-
zation and relatedness, it is interpreted here as a
map of the phenotypic relationships between
diverse cell types and is not meant to imply a
developmental hierarchy . Indeed, even measur-
ing a large number of cells in a single tissue
will fail to capture some developmental tran-
sitions, including (i) rapid activation (release of
cytoplasmically sequestered receptors) (29); (ii)
uneven surface marker partitioning during asym-
metric cell division (30); and (iii) organ -specific
development outside the assayed organ (matura-
www.sciencemag.org SCIENCE VOL 332 6 MAY 2011
689
RESEARCH ARTICLE
on May 14, 2011www.sciencemag.orgDownloaded from
Page 3
tion of T cells in the thymus). In this bone mar-
row data set, several well-defined cell types (such
as T, NK, B, and monocyte) provide landmarks
for the organization of the tree and give context to
the nodes encompassing transitional and less-
understood cell types. Ultimately , this ap pr oa ch
enabled visualization of 34-dimensional bone mar-
row data in an intuitive graphical format. Al-
though the algorithm over-segregated some cell
types into redundant contiguous clusters, this ap-
proach has several advantages that complement ed
the complexity of this data set: (i) increased resolu-
tion captured unexpected and transitional cell
types that escape standard classification strategies;
(ii) Unsupervised analysis helped overcome the
bias of subjective gating; and (iii) n-dimensional
algorithms leveraged the multi-parameter mass
cytometry data to define cell types on the basis of
previously unappre ciated, subtle differences in sur-
face expression. Although we used stochastically
selected seed cells to initiate the tree genera-
tion along with local similarity clustering and
minimum-spanning trees, the approach is ame-
nable to incorporation of other more determi-
nistic partitioning approaches that might allow
for other standardized tree structure formation.
Ex vivo analysis of healthy human bone mar-
row signaling. Historically, by detailing im-
mune functions in vivo and in vitro a model of
specialized cell types in immunology and hema-
topoiesis were mapped primarily on the basis of
expressed cell surface antigensmany of which
were codified by using single-cell analysis and
fluorescence-based cytometry (3134 ). Because
cell-surface proteins represent only a small pro-
portion of the repertoire of gene products gov-
erning cell behavior, intracellular proteins (33)
are also critical in defining cell types. Because
surface and intracellular molecules work to-
gether in concert to support different cellular
roles, it might be expected that proteins govern-
ing specialized immunological cell functions
(T cell receptor, B cell receptor, or cytokine re-
ceptors) are modulated in a coordinated manner
as cells transit developmental pathways from
stem cell precursors to differentiated endpoints.
We monitored 13 surface markers to iden-
tify immune cell types and 18 intracellular epi-
topes in order to interrogate intracellular signaling
biology in healthy human bone marrow. We
examined the signaling dynamics of these 18 in-
tracellular markers in response to 13 ex vivo stim-
ulation conditions (such as IL-7 or GM-CSF),
including those shown to have prognostic value
in leukemia, lymphoma, and myeloproliferative
disorders [such as granulocyte colony stimulat-
ingfactor(G-CSF)](10, 3537 ). Cell populations
were first defined on the basis of conventional sur-
face expression gates, ultimately identifying 24 im-
munological populations in human bone marrow
Fig. 2. SPADE links re-
lated immune cell types
in a multidimensional con-
tinuum of marker expres-
sion. (A)Summaryof
SPADE analysis. Single-
cell data are sampled in
a density-dependent fash-
ionsoastoreducethe
total cell count while main-
taining representation
of all cell phenotypes.
Neighboring cells are then
grouped by unsupervised
hierarchical clustering. Re-
sulting nodes (defined as
those cells within a bound-
ary of an n-dimensional
hull) are then linked by
a minimum-spanning tree,
which is flattened for 2D
display. (B) Immunopheno-
typic progression in healthy
human bone marrow. A
tree plot was constructed
by using 13 cell-surface
antigens in healthy hu-
man bone marrow. 18
additional intracellular
parameters were acquired
concurrently but excluded
from t ree construction.
Thesizeofeachcirclein
the tree indicates rela-
tive frequency of cells
that fall within the 13-
dimensional confines of
the node boundaries. Node color is scaled to the median intensity of marker
expression of the cells within each node, expressed as a percentage of the maximum
value in the data set (CD45RA is shown). Putative cell populations were annotated
manually (table S2) and are represented by colored lines enci rcling sets of nodes
that have CD marker expression emblematic of the indicated subset designations.
(C) Overlaid expression patterns of CD3, CD8, and CD4. Three markers, along with
CD45RA (B), were used in clustering that helped define T cell lineages. Color scale is
as described in (B). (D) Overlaid expression patterns of CD19, CD20, and CD38.
Three markers were used in clustering that helped define B cell lineages. Color scale
is as described in (B). (E) Overlaid expression patterns of CD34, CD123, and CD33.
Three markers were used in clustering that helped define myeloid and progenitor
cell lineages. Color scale is as described in (B). (F) Overlaid expression of com-
plementary surface markers from a staining panel with 18 additional surface
markers (fig. S4) by using the 13 core surface markers as landmarks (21). Overlaid
expression patterns are shown for eight complementary surface markers that helped
to further define the myeloid (CD13, CD14, and CD15), B cell (CD10), and NK/T cell
(CD7, CD56, CD161, and CD16) portions of the SPADE representation. These
markers were not used for tree construction. Color scale is as described in (B).
CD20CD19 CD38
MST Using
n Dimensional
Coordinates
A
CD33
CD123
CD34
CD161
BC
D
0100
Expression
% Maximum
CD15
CD16
CD7
CD13 CD14
CD45RA
CD3 CD8
CD4
EF
CD10
CD56
Parameter 1
Parameter 2
Parameter 3
Density-
Dependent
Downsampling
Local
Clustering in
n
Dimensions
Memory CD8+ TNK
Pre-B II
Plasma cell
Pre-B I
MPP
Pro-B
HSC
Monocyte
CD38
mid
CD3- platelet
CD38- CD3- platelet
CD38
mid
CD3
mid
platelet
Erythrocyte
Erythroblast
Myelocyte
Plasmacytoid DC
Promyelocyte
Pre-DC
GMP
Pro-monocyte
CMP
Monoblast
NKT
Naive CD8+ T
Naive CD4+ T
Memory CD4+ T
IL-3Rα+ mature B
Mature B
Immature B
6 MAY 2011 VOL 332 SCIENCE www.sciencemag.org690
RESEARCH ARTICLE
on May 14, 2011www.sciencemag.orgDownloaded from
Page 4
(fig. S5). The induced intracellular signali n g re-
sponses (changes in phosphorylation state) in
these populations, as compared with those of
an untreated control, were summarized as a hea t -
map (Fig. 3A). Unsupervised, hierarchical clustering
of the phosphorylation responses allowed distinc-
tion of biologically related cell types (T cell subsets)
by their signaling behavior alone, demonstrating
that signaling capacities are closely tied to cel-
lular lineage (fig. S6). Several canonical signaling
responses that mapped to manually determined cell
typesareshowninFig.3B.Theseextremelyspe-
cialized responses, such as the tight restriction of
IL-7mediated pSTAT5 responsiveness in T cells
(Fig. 3B, arrow 4) (38) or lipopolysaccharide (LPS)
stimulated phosphorylation of the mitogen-activated
protein kinase (MAPK) p38 ( p-p38) responsiveness
in monocytes (Fig. 3B, arrow 5) (39), suggest the
existence of correlations between signaling events
and surface markerde fined boundaries, thus pre-
senting an opportunity to establish a unified view
of immune signaling during hematopoiesis.
W ith ~10
4
signaling observations (Fig. 3A
and fig. S10A) for each replicate bone marrow ,
it was necessary to filter the data set in order to
arrive at the most significant and potentially novel
observations . Using a o ne-sample t-test, over 500
observations were observed with a Bon ferroni-
adjusted significance of P < 0.05 in each replicate
bone marrow for a total of 860 unique responses
(fig. S7 and table S3). Of the 248 observations
overlapping between patient marrows, 28 be-
longed exclusively to cells residing in the hu-
man hematopoietic progenitor cell compartment
[hematopoietic stem cells (HSCs), multipotent
progenitors (MPPs), granulocyte/macrophage pro-
genitors (GMPs), and megakaryocyte-erythroid
progenitors (MEPs)], including G-CSF induc-
tion of pSTAT3 in the most primitive cell types,
HSC and MPP (40). This same signaling behav-
ior correlated with negative prognosis in acute
myeloid leukemia (10), suggesting that, as in the
case of other malignancies, there may be a selec-
tive advantage for cells to mimic the properties of
their most primitive counterparts.
For a more objective and fine-grained view of
these cell typespecific responses, free of the
biases of conventional 1D and 2D surface marker
categorization, we overlaid the signaling behav-
ior of the 18 functional epitopes on the tree structure
using a similar approach as described for the im-
munophenotype staining panel (Fig. 2), allowing
the intracellular signaling status to be vis u al iz e d
on the previously annotated tree structure (Fig. 3C).
Nodes were colored according to the magnitude of
the difference in their median responses relative to
the untreated control. This effectively eliminated
the subjectivity of manual classification and im-
proved the resolution of the heatmap (Fig. 3A),
separating the 24 manually assigned cell types
into 282 logically connected node s of phen o ty p-
ically distinct, but locally similar, cell clusters.
The stimuli that corresponded closely with cell
types identified manually in the heatmap also ex-
hibited appropriately specific responses when over-
laid on the tree structurespecifical ly , IL-7/pST AT5
in T cells, B cell receptor (BCR)/phosphorylated
B cell linker protein ( pBLNK) exclusively in im-
mature and mature B cells, and LPS/p-p38 re-
stricted to the monocyte compartments (Fig. 3C),
with the latter corresponding to the ex