PPT

Report
Microarray Gene Expression Analysis
Differential expression, clustering,
networks, and functional enrichment
STEMREM 201 Fall 2012
Aaron Newman, Ph.D.
10/17/12
A genomics approach to biology involves…
 Finding significant patterns in high-throughput data
 Interpreting these patterns in the context of prior knowledge
 Generating new hypotheses and predictions
• A plethora of techniques and tools exist…
• My goal is to introduce some practical, powerful, and freely
available methods for gene expression analysis
Example workflow
Gene expression profiling
DEGs
Cluster
analysis
Functional
enrichment
Network
analysis
Biological
meaning
GSEA
Materials and Methods
GSE22651
ESCs and differentiated cells
Excel
&
GenePattern
Hierarchical
&
AutoSOME
DAVID
Toppfun
STRING
&
Cytoscape
Biological
meaning
GSEA
Microarray normalization
•
Probe/array-level normalization (reduce inter-array technical errors and
noise)
– Raw CEL files  merged and normalized text file
• Robust Multi-Chip Averaging (RMA) for Affymetrix arrays
–
Affymetrix Gene Expression Console software
• Quantile Normalization for Agilent, Illumina, and others.
– Sets all arrays to the same distribution (mean, median, sd, etc.)
•
Analysis-level normalization
– Log2 Transformation
•
Improves statistical properties for analysis (log-normal)
– Median/mean centering
•
Useful for reducing impact of transcript
abundance on identifying/visualizing
co-expressed genes
center
– Unit variance
•
Standardize each column (array) to mean of 0 and standard deviation of 1
Differentially expressed genes
(DEGs)
• Goal: given known phenotypes, determine which genes
exhibit significant differential expression.
• Many genes will be tested; multiple hypothesis testing should
be performed to control the false discovery rate (FDR).
– Bonferroni (α / n)
– Benjamini-Hochberg (largest k s.t. P(k) ≤ (αk) / n)
– Storey Q-value (p-value specific FDR)
• Q-value software - http://genomics.princeton.edu/storeylab/qvalue/
• Sample permutations can improve p-value accuracy
– *Only apply with ≥ 10 samples per class
– Implemented in GenePattern (ComparativeMarkerSelection
module), SAM
Supervised DEG identification tools
(i.e., classes are known)
• Likely familiar
– Excel (T-test) coupled with FDR assessment
• Watch for conversion to dates! (e.g., MARCH6  3/6)
• Basic to Intermediate
– Statistical Analysis of Microarrays (Windows only)
– GenePattern (Broad Institute)
– p.adjust (R)
• Advanced
– Bioconductor packages in R (e.g., limma)
Tutorial 1:
Identifying DEGs in Excel using FDR cutoff of 1% (BH method)
Open “DEG_example_large.txt” in Excel (2 classes, 25151 genes).
In column L, use T-test function to test for significant differential
expression between ESCs and non-ESCs.
1)
2)
•
•
3)
4)
5)
=TTEST(“ESCs”, “non-ESCs”, 2, 3)
2-tailed, unpaired with unequal variance
Sort p-values in column L in ascending order.
In column M, input p-value rank, going from 1,2,3…n.
Input following formula into column N to test for FDR of 1%
•
•
•
6)
In general: = (0.01 * rank) / n
In our case: = (0.01 * M2) / 25151
Autocomplete column N
In column O, test for p-values that do not exceed FDR of 1%
•
= if(L2 <= N2, 1, 0)
That’s it! All genes with a 1 in column O are significant at an FDR
of 1%.
7)
•
Should be 3237 significant genes.
Cleaning up result and creating a heat map
1)
Calculate log2 fold change in column P
•
•
= AVERAGE(B2:F2) – AVERAGE(G2:K2)
Autocomplete
Use filter to isolate significant genes with absolute fold change ≥ 5
Copy and paste into new tab
Sort by fold change in descending order
Save as new file (e.g., DEG_example_sorted_fdr01_fold5.txt)
Center genes in Cluster 3.0
2)
3)
4)
5)
6)
•
•
•
•
7)
Open file
Go to “Adjust Data”
Check “Center genes” and select “Median”
Press “Apply”, then save file (e.g.
DEG_example_sorted_fdr01_fold5_medcen.txt)
Open in Java TreeView.
•
•
To customize heat map display and text, use “Settings>Pixel Settings”, and
“Settings>Font Settings”.
Export heat map image using “Export>Save Tree Image”.
Expected result
FDR ≤ 1%
Fold change ≥ abs(25)
= 79 genes
Tutorial 2:
Identifying DEGs in GenePattern
1)
2)
3)
4)
Create account (if you have not already) and log in to GenePattern
Select “Differential Expression Analysis”.
Skip to ComparativeMarkerSelection (step 2) and click “Open module”.
Enter following fields:
•
•
•
5)
6)
7)
8)
9)
input* file = GSEA_example_expression_large.gct
cls* file = GSEA_example_classes.cls
number of permutations* = 0
Select “Run”
When processing is finished, select “ComparativeMarkerSelectionViewer”
from the pulldown menu next to
“GSEA_example_expression_large.comp.marker.odf”
Select “Run”
Select “Open Visualizer” and “Allow”
A table will appear showing all genes and various statistics, including:
1)
2)
3)
4)
Benjamini_Hochberg corrected p-value; FDR(BH)
Storey q-value; Q Value
Bonferroni p-value
Fold change
Extracting data and displaying as a heat map
1)
2)
3)
4)
5)
Pipe .odf file to “ExtractComparativeMarkerResults”
module.
Select genes with BH FDR ≤ 1%.
Filtered data are available as a new .gct file.
To display a heat map, pipe .gct file to “HeatMapViewer”
module.
Select “Run”
Example workflow
Gene expression profiling
DEGs
Cluster
analysis
Functional
enrichment
Network
analysis
Biological
meaning
GSEA
The “Clustering Problem” for Large Data Sets
Widespread Importance
– Genomics
– Phylogenetics
– Disease
– Galaxy Clusters
– etc., etc., etc.
Common clustering methods
Toy data set
K-means
Hierarchical (Eisen)
SOM
Figure 1. D’haeseleer, Nat Biotechnol. 2005
Feature Comparison
Method
Handle
Diverse
Large
Cluster
Datasets Shapes
√
Hierarchical
K-Means
√
SelfOrganizing
Map (SOM)
√
Detect
Cluster
Number
Identify
Outliers
Low
Output
Variance
√*
√
√*
Tutorial 3: Hierarchical Clustering
1)
Open “GSE22651_filtered.txt” in Cluster 3.0
1)
Normalize data (Adjust Data tab)
• Check “Log transform data”
• Check “Center genes” and select “Median”
• Press “Apply Button”
2)
Cluster data (Hierarchical tab)
• Check “Cluster” under Genes and under Arrays
• Leave “Similarity Metric” at Uncentered Correlation
Press “Centroid linkage” under “Cluster method”
3)
•
•
•
•
4)
Centroid = distance between cluster centers
Single = closest distance between clusters
Complete = farthest distance between clusters
Average = mean of all pairwise distances between clusters
Open clustered data table file (*.cdt) in Java TreeView.
Java TreeView
Navigate the
cluster tree
to highlight
genes with
distinct
expression
patterns in
particular
samples
Export>”Save
List” to copy
or save gene
lists of
interest for
further
analysis.
Expected result
Genes
Samples
Automatic clustering of
Self-Organizing Map Ensembles
AutoSOME
Serial application of
- SOM
- Density Equalization
- Minimum Spanning Tree
- Ensemble Averaging
Newman and Cooper (2010) BMC Bioinformatics, 11:117
AutoSOME
Method
Handle
Diverse
Large
Cluster
Datasets Shapes
Detect
Cluster
Number
√
Hierarchical
Identify
Outliers
Low
Output
Variance
√*
√
K-Means
√
√*
SelfOrganizing
Map (SOM)
√
Affinity
Propagation
√*
Spectral
Clustering
√*
√
√
nNMF
√
√
√
AutoSOME
√
√
√
√
√
√
√
√
AutoSOME Webstart
http://jimcooperlab.mcdb.ucsb.edu/autosome
Tutorial 4:
Identifying discrete clusters without prior knowledge of cluster number
1)
2)
3)
4)
Launch the AutoSOME GUI via the large launch button.
Open “GSE22651_filtered.txt”
Skip filtering
Show “Basic Fields”
•
Set p-value = 0.05
Show “Input Adjustment”
5)
•
Check Log2 Scaling, Unit Variance, Median Center, and Sum Squares = 1
Press the large “Run” button on the left.
From the menubar, select View>heat map> green red.
Select cluster 1 in the cluster list.
The data are rendered as a normalized heat map. To change the
display, go to View>settings>image settings.
6)
7)
8)
9)
•
•
Under the Normalization tab, check “Display Original Data”, “Log2 Scaling”,
and “Median Center”.
Check “Manually adjust range for contrast” and set minimum to -2 and
maximum to 2. Press “Update” (lower left corner).
AutoSOME continued…
1)
2)
3)
Select clusters 1 to 5 (hold down shift). Right click mouse in heat map window to resize.
Set “Zoom Factor” to 40 and Press “Save”.
See website for further tutorials and documentation.
Representative heat map
View>heat map>rainbow
Example workflow
Gene expression profiling
DEGs
Cluster
analysis
Functional
enrichment
Network
analysis
Biological
meaning
GSEA
How to interpret clustering results
• Gene set functional annotation
– DAVID
– Toppfun
– MSigDB
http://david.abcc.ncifcrf.gov/
http://toppgene.cchmc.org/enrichment.jsp
http://www.broadinstitute.org/gsea/msigdb/
• Network analysis
– STRING
http://string-db.org/
Gene sets: where do they come from?
• Gene ontology
– An attempt to semantically organize genes and their functional
relationships.
– Data are arranged in a graph structure, from broad to specific
– Ontologies:
• Biological process (BP): series of ordered events
• Molecular function (MF): activities that occur at the
molecular level
• Cellular component (CC): part of a cell
• Biocarta/KEGG pathways (curated wiring diagrams)
• High-throughput studies
e.g., MSigDB
Tutorial 5: Functional annotation of a large cluster using DAVID
http://david.abcc.ncifcrf.gov/
1)
In AutoSOME, find the cluster of genes with
higher expression in stem cells.
2)
Go to View>raw data. Highlight all genes in
the cluster and copy the list.
3)
Open the DAVID homepage and press “Start
Analysis”
Select “Upload” tab and paste in gene list.
Under “Select Identifier,” select
“OFFICIAL_GENE_SYMBOL”.
Select “Gene List” for “List Type”.
Submit the list.
Press OK at multi-species warning message.
Select “Homo sapiens” and press “Select
Species”.
Select “Functional Annotation Tool”
Press “Functional Annotation Clustering”
button.
4)
5)
6)
7)
8)
9)
10)
11)
DAVID Output
Similar gene sets are clustered together, eliminating redundancy and facilitating
interpretation
Gene sets
Pathways
Tutorial 6: Protein-protein associations with STRING
Top 50 non-ESC genes
Protein-protein association network among top 50
non-ESC genes
Evidence
types
Top 100 non-ESC genes
Example workflow
Gene expression profiling
DEGs
Cluster
analysis
Functional
enrichment
Network
analysis
Biological
meaning
GSEA
Gene set enrichment in a ranked list
• Gene Set Enrichment Analysis (GSEA)
• “Threshold-less”
– Arbitrary DEG cutoffs are avoided
• Two modes of operation to rank input genes:
– Rank by differential expression between phenotypes
• Default metric is signal to noise, defined as:
(avg[class 1] – avg[class 2]) / (sd[class 1] + sd[class 2])
– Pre-ranked according to user-defined criteria
• Evaluates statistical bias in the distribution of each defined gene set
over the list of ranked input genes.
Input format
• Input:
– Expression data (gene cluster text file format, *.gct)
– Classes (*.cls)
• File extensions matter!
– If using Notepad in Windows, set Save as type: to “All Files
(*.*)
– If using TextEdit in Mac, go to Preferences > Open and Save >
and uncheck ‘Add “.txt” extension to plain text files’.
• Formatting instructions:
http://www.broadinstitute.org/cancer/software/gsea/wiki/index.php/Data_for
mats
Tutorial 7: GSEA
 Input files: GSEA_example_expression_large.gct, GSEA_example_classes.cls
Load data
Run GSEA
Output
Summary
Gene expression profiling
DEGs
Cluster
analysis
Functional
enrichment
Network
analysis
Biological
meaning
GSEA

similar documents