Download presentation
Published byChad Mosley Modified over 9 years ago
1
Tools for visualizing high dimensional single cell data and inferring cellular hierarchy
Sylvia K. Plevritis, PhD Professor Department of Radiology and (by courtesy) Management Science and Engineering Stanford University School of Medicine
2
Clustering versus Ordering
Most microarray analyses are focus on what is the difference between A and B? We want to know: Is it possible that A becomes B (or visa versa)? If so, what are the molecular drivers of this process? B A All these existing methods are essentially asking the same question: What is the difference between A and B. In this talk, we are going to ask a different question. How did A become B? By that, I mean Does there exist any gradual/progressive changing pattern among individual samples? Does there exist such a progression (green line) that describes: (1) the progression among individual samples within group A, (2) the progression within group B, and (3) the transition in-between. If we can identify such a progressive pattern, this pattern tells us the trajectory of biological progression, and the important transition point. This can be applied to developmental biology, disease progression, drug response. B A
3
Inferring Order Given high throughput datasets, we want to identify:
(1) the ordering among individual samples (2) which markers identify the ordering Here is our goal, Given a microarray data matrix, we want to identify two things: (1) (2)
4
Sample Progression Discovery (SPD)
Qiu et al., Discovering Biological Progression Underlying Microarray Samples, PLoS Computational Biology, 2011.
5
SPD on B-cell differentiation
7 HSC 7 CLP 7 proB 7 preB 7 IM 5 M (Naïve B, CB, CC, Memory B, CD19+) Microarray expression data is obtained from the Weissman Lab.
6
SPD on B-cell differentiation
7
SPD on B-cell differentiation
Genes in selected modules were specific to B-cell differentiation and included CD19, CD20, CD79 as well as master transcription factors including PAX5 and SP140. There was also enrichment of genes in the BCR pathway. Explain the genes that are responsible for this ordering.
8
SPD infers Individual Tumor Plasticity in TCGA Breast Cancer Data
Color coded by PAM50 Score Enrichment of Mammary Gland Development, Mesenchymal Cells
9
SPD Graph Color-Coded by Genes Associated with EMT
Breast Cancer SPD graph with the samples colored with the expression of the module that contained genes involved in mesenchymal processes.
10
Visualizing and Ordering High Dimensional Single Cell Data
11
Single cell mass cytometry
Garry Nolan, PhD CyTOF = Cytometer + Elemental Mass Spectrometer CyTOF Data Sample: normal human bone marrow 31 Proteins measured on single cells 13 core surface markers 18 function markers
12
Qiu et al., Nature Biotechnology, 2011.
SPADE: Spanning-tree Progression Analysis of Density-normalized Events Qiu et al., Nature Biotechnology, 2011. Anchang et al, Nature Protocols, (accepted).
13
Viewing all the surface markers …
14
Manually identifying the populations …
15
Pooling samples …
16
Analyzing a non-branching process …
17
(a) Normal Bone Marrow (b) ALL
Comparison to other visualization algorithms … (a) Normal Bone Marrow (b) ALL viSNE : Amir et al., Nature Biotechnology 2013. ACCENSE: Shekhar et al., PNAS 2014.
18
Integration of SPADE and t-SNE to create “SPADE FOREST”…
19
Multi-target drug combinations derived from single drug effects measured at the level of single cells
20
Overview Intratumor heterogeneity is modeled at the level of the single cell Single drug response at the level of single cell is measured by changes in protein expression using CyToF Analysis of drug response are performed on clusters of similar cells Drug combinations are derived from single drug response each cell cluster individually then combined through a mixture model
21
Intratumoral Heterogeneity
22
Outline Collect perturbation single data drug response data using CyTOF Identify homogeneous cell types based on surface markers Build a “nested effects” graphical drug model in each cell type based on intracellular change; estimate a “mixture nested effects” Determine a scoring function of optimizing drug combinations identify homogeneous cell types based on surface markers build a “nested effects” graphical model of a drug response in each cell type based on intracellular change; estimate a “mixture nested effects” determine a scoring function of optimizing drug combinations 1 Decision tree
23
Outline Collect single data perturbation data using CyTOF
Identify homogeneous cell types based on surface markers Build a “nested effects” graphical drug model in each cell type based on intracellular change; estimate a “mixture nested effects” Determine a scoring function of optimizing drug combinations identify homogeneous cell types based on surface markers build a “nested effects” graphical model of a drug response in each cell type based on intracellular change; estimate a “mixture nested effects” determine a scoring function of optimizing drug combinations 1 Decision tree
24
Perturbation CyTOF Profiling of Functional & Surface Phenotypes in Healthy Bone Marrow and Pediatric AML p4EBP1 pPLCgamma2 pAkt pRb pAMPK pS6 pCbl pSTAT1 pCREB pSTAT3 pERK1/2 pSTAT5 p-p38 MAPK cleaved Casp3 pSyk 15 Functional Markers CD3 CD41 CD7 CD44 CD11b CD45 CD15 CD47 CD19 CD64 CD33 CD117 CD34 CD123 CD38 HLADR 16 Surface Markers Staining panel (31 Abs) No inhibitor Basal (unstim.) AICAR Flt3 ligand G-CSF GM-CSF IFNα IFNϒ IL-3 IL-6 IL-10 IL-27 PMA + ionomycin PVO4 SCF TNFα TPO Perturbations (19) PI3K/mTOR inhib. Inhibitor alone PMA/ionomycin Healthy donor marrow (7) Diagnosis AML marrow (18) Matched relapse AML marrow (3) Courtesy of Nolan Lab.
25
Outline Collect single data perturbation data using CyTOF
Identify homogeneous cell types based on surface markers Build a “nested effects” graphical drug model in each cell type based on intracellular change; estimate a “mixture nested effects” Determine a scoring function of optimizing drug combinations identify homogeneous cell types based on surface markers build a “nested effects” graphical model of a drug response in each cell type based on intracellular change; estimate a “mixture nested effects” determine a scoring function of optimizing drug combinations 1 Decision tree
26
CCAST: Classification, Clustering and Sorting Tree
Anchang et al., PLoS Computational Biology, 2014.
27
CCAST identifies more homogeneous B-cell subpopulations
Anchang et al., PLoS Computational Biology, 2014.
28
Outline Collect single data drug response data using CyTOF
Identify homogeneous cell types based on surface markers Build a “nested effects” graphical drug model in each cell type based on intracellular change; estimate a “mixture nested effects” Determine a scoring function of optimizing drug combinations identify homogeneous cell types based on surface markers build a “nested effects” graphical model of a drug response in each cell type based on intracellular change; estimate a “mixture nested effects” determine a scoring function of optimizing drug combinations 1 Decision tree
29
Graphical Models of Nested Drug Effects
An effect represents a change in protein marker following drug intervention Different drug effect subsets can be observed In a graphical model of the drugs D1 and D2, an edge from D1 to D2 indicaes that the effects of D2 are nested in the effects of D1 Node represent the drugs An edge from D1 to D2 indicaes that the effects of D2 are nested in the effects of D1 An effect represents a change in signal from control drug intervention Different drug effect subsets can be observed
30
Objective Function of DRUG NEM
Given a drug response single cell data from n drugs and m targets, identify an optimal drug regimen as minimum number of drugs that maximum number of markers effected. 1 Decision tree
31
Multi-drug experiment on HeLa cells under TRAlL stimulation
Study design Stimulation: TRAIL (Base line treatment) Inhibitors: JNK 1, GDC, GSK, SB Cell states : Apoptotic-like and survivor-like Intracellular markers: 1 Decision tree
32
MNEMs indicate pP38 MAPK (SB) and Mek (GSK) inhibitors are important candidates for combination therapy We would SB and GSK to top the list of best 2 drug combinations 1 Decision tree
33
Summary SPD infers hierarchical ordering of tumors based on gene experssion SPADE infers hierarchical ordering among single cells based on CyTOF data DRUGMNEM generates drug combination hypotheses using single drug information based on measurements of intratumor heterogeneity
34
Acknowledgements Benedict Anchang Peng Qiu Kara Davis Harris Feinberg
Sean Bendall Robert Tibshirani Garry Nolan NCI Integrative Cancer Biology Program (ICBP)
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.