Erin Simonds, PhD W. A. Weiss Lab, UCSF
May 23, 2013
Tools and approaches for mass cytometry
data analysis
© 2011 J. Simonds
Biaxial plots are not a scalable solution
Parameters: 4 8 14 32
Plots: 6 28 91 496 2010 2000 1989 1983
How can we explore in 30+ dimensions?
The cytometrist’s toolbox
Radar
FLAME
PCA
Heatmap
Biaxial plot Histogram
FLOCK SamSPECTRAL FlowMerge FlowClust
Gemstone SPADE viSNE
Transform raw data
3D plot
Reduce dimensionality
Summarize statistics
Identify clusters
Wanderlust
Box plot
Vetting the CyTOF: In a fair fight, nearly identical data
Experimental design 7 Surface Markers (CD 3, 4, 8, 20, 33, 45RA, 56) 2 Functional Markers (pSTAT3 & 5)
Bendall, Simonds, et al. Science, May 2011
The cytometrist’s toolbox
Radar
FLAME
PCA
Heatmap
Biaxial plot Histogram
FLOCK SamSPECTRAL FlowMerge FlowClust
Gemstone SPADE viSNE
Transform raw data
3D plot
Reduce dimensionality
Summarize statistics
Identify clusters
Wanderlust
Box plot
SPADE: Spanning-tree Progression Analysis of Density-normalized Events
Qiu et al., Nature Biotechnology, Oct. 2011
Marker 1
Mar
ker 1
Marker 2
Marker 2
SPADE
CD45 CD3 CD4 CD8 CD34 CD19 CD20 CD33 CD11b CD123
SPADE projects bone marrow as a continuum of phenotypes
0 100 Intensity (%max)
Bendall, Simonds, et al. Science, May 2011
pSTAT5
0 100 Intensity (%max)
Rediscovery of canonical signaling pathways
Basal IL-7
Bendall, Simonds, et al. Science, May 2011
+ Dasatinib Abl / Src-family kinase inhibitor
Dasatinib differentially affects immune cell subsets
Bendall, Simonds, et al. Science, May 2011
IL7 (pSTAT5)
PVO4 (pSTAT5)
The cytometrist’s toolbox
Transform raw data
Reduce dimensionality
Summarize statistics
Identify clusters
Radar
FLAME
PCA
Heatmap
Biaxial plot Histogram
FLOCK SamSPECTRAL FlowMerge FlowClust
Gemstone SPADE viSNE
3D plot
Wanderlust
Box plot
viSNE preserves high-dimensional, non-linear relationships
Raw data (in 3D) viSNE map (in 2D)
z
y x
viSNE-1
viS
NE
-2
Amir et al., Nature Biotechnol. ePub 19 May 2013
viSNE maps healthy marrow as discrete populations
Amir et al., Nature Biotechnol. ePub 19 May 2013
CD4+ T
CD11b-hi monocyte
CD11b-lo monocyte
NK
CD20+ B
CD8+ T
CD20- B
Amir et al., Nature Biotechnol. ePub 19 May 2013
viSNE can help detect rare subpopulations
ALL cells
Barcode A: Healthy bone marrow cells Barcode B: ALL cells spiked into sample at 0.25% frequency (1/ 400 cells)
Amir et al., Nature Biotechnol. ePub 19 May 2013
Same data, different dimensionality reduction algorithms
viSNE is still stochastic, but much more reproducible
Amir et al., Nature Biotechnol. ePub 19 May 2013
SPADE’s stochasticity causes difference between runs
So why should I ever use SPADE? à Clustering allows comparison between samples
The cytometrist’s toolbox
Transform raw data
Reduce dimensionality
Summarize statistics
Identify clusters
Radar
FLAME
PCA
Heatmap
Biaxial plot Histogram
FLOCK SamSPECTRAL FlowMerge FlowClust
Gemstone SPADE viSNE
3D plot
Wanderlust
Box plot
pSTA
T5
pSTAT3
Goal: Characterize the GCSF-responsive subpopulation
G-CSF
pSTAT3 pS
TAT5
Who are these cells?
Are they the same in every
patient?
Surface marker profiling of G-CSF responders in AML
with Faye Hsu, Yury Goltsev
pSTAT3
pSTA
T5
CD14 CD32 CD36 CD86 CD93 CD11a CD1d HLA-‐DR
CD46 CD69 CD117 CD133 CD155 CD321
CD13 CD15 CD33 CD34 CD38 CD44 CD45
Non- responders
+ markers from literature
CD47 CD64 CD114 CD123 CD11b CXCR4 TIM3
Responders
Differential gene expression
Screen pediatric AML samples (18 diagnosis, 3 relapse, 7 healthy)
LSCs
Gate G-CSF responders
Identify enriched markers
pSTAT3
pSTA
T5
Basal G-CSF
pSTAT3
pSTA
T5
CD46
Leukemic GCSF responders have distinct surface profiles
with Rob Bruggner & Sean Bendall
Growth factor
receptors
Universal markers
CD34
LSC markers
pSTAT3
pSTA
T5
Healthy
The cytometrist’s toolbox
Transform raw data
Reduce dimensionality
Summarize statistics
Identify clusters
Radar
FLAME
PCA
Heatmap
Biaxial plot Histogram
FLOCK SamSPECTRAL FlowMerge FlowClust
Gemstone SPADE viSNE
3D plot
Wanderlust
Box plot
Human B cell development is less understood than in the mouse
CD34
CD10
CD24 IL7R
CD38
CD179a
Surface Markers
Intracellular Markers
Rag TdT
IgH Genes
CD19
Adapted – Cobaleda, BioEssays 2009
DJH
VDJH
Maturity Trajectory
The predicted trajectory ‘rediscovers’ human B cell development
CD34
TDT
“Wanderlust” algorithm: El-ad Amir, Dana Pe’er (Columbia)
CD24
CD34
CD
38
Following the predicted B cell trajectory
1
• Fluorescent panel design à Sorting à qPCR
• VDJ rearrangement confirms the trajectory: (unrearranged) 1 à 2 à 3 à 4 à 5 (rearranged)
Sean Bendall, Kara Davis
CD24
TdT
4
2 5
3
New insights and resolution in human B cell development
CD34
CD10
CD24 IL7R
CD38
CD179a
Surface Markers
Cytoplasmic Markers
Rag TdT
IgH Genes
CD19
DJH VDJH
I IIa IIb III IV Va Vb
✓
✓
CD179b Ki67
pSTAT5
Sean Bendall, Kara Davis
Summary
Radar
FLAME
PCA
Heatmap
Biaxial plot Histogram
FLOCK SamSPECTRAL FlowMerge FlowClust
Gemstone SPADE viSNE
Transform raw data
3D plot
Reduce dimensionality
Summarize statistics
Identify clusters
Wanderlust
Box plot
Sean Bendall Kara Davis Harris Fienberg Astraea Jager Rob Bruggner Rachel Finck
stanford.edu/group/nolan
Mount Sinai Michael Linderman
Columbia El-Ad Amir Jacob Levine Dana Pe’er
St. Jude Children’s Amanda Gedman Ina Radtke James Downing
Stanford Peng Qiu (à MD Anderson) Sylvia Plevritis
Prof. Garry P. Nolan Bernd Bodenmiller (à U. Zurich) Eli Zunder Greg Behbehani Wendy Fantl and the entire Nolan lab