1. Microarray Gene Expression Analysis
23/03/2009
Daniele Merico
PhD, Molecular and Cellular Biology
Bader Lab & Emili Lab

2. Gene expression analysis: general workflow
Define the experimental design
Collect the biological samples
Generate the expression data
Identify the
Differential Genes
Identify the
Functional Groups

3. Identify the Functional Groups
Different Strategies
GENE SETS
PATHWAYS
NETWORKS
Spindle
Gene.1
Gene.2
Gene.3
P53 signaling
Gene.2
Gene.4
Gene.5
Score the set depending on the gene expression of its member genes
Just visual, or
Score the pathways exploiting gene expression and topology
Identify sub-networks (i.e. modules) satisfying some joint gene expression and topology requirement

4. A brief history of life
microarrays
About 5 min.

5. Microarray Chronology
Number of PubMed publications by year
Using a query containing keywords such as microarray, transcriptomics, etc..

6. Microarray Chronology
First Microarray Publication [1]
45 Arabidopsis genes
[1] Schena M, Shalon D, Davis RW, Brown PO.; Quantitative monitoring of gene expression patterns with a complementary DNA microarray.; Science Oct 20;270(5235):

7. Microarray Chronology
Full Yeast Genome on microarray [2]
[2] Lashkari DA, DeRisi JL, McCusker JH, Namath AF, Gentile C, Hwang SY, Brown PO, Davis RW. Yeast microarrays for genome wide parallel genetic and gene expression analysis. Proc Natl Acad Sci U S A Nov 25;94(24):

8. Microarray Chronology
Gene Ontology Consortium.
Hierarchical Clustering and heat-maps [3]
[3] M. B. Eisen, P. T. Spellman, P. O. Brown and D. Botstein, Cluster analysis and display of genome-wide expression patterns. Proc. Natl. Acad. Sci. USA 95 (1998), pp –14868.

9. Microarray Chronology
Gene Ontology enrichment, (hypergeometric)

10. Microarray Chronology
Gene expression profiling on interaction networks [4]
[4] Discovering regulatory and signalling circuits in molecular interaction networks. Ideker T, Ozier O, Schwikowski B, Siegel AF. Bioinformatics. 2002;18 Suppl 1:S

11. Microarray Chronology
Full Human Genome on microarray
Affymetrix HGU-133 plus 2.0

12. Microarray Chronology
GSEA Enrichment [5]
Gene Ontology, Pathways, other gene-sets
[5] Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, Mesirov JP. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles.
Proc Natl Acad Sci U S A Oct 25;102(43):

13. Gene expression analysis: general workflow
Define the experimental design
Generate the expression signals
Explorative Analysis
and Pre-processing
Select Diff. Genes
Group into Clusters
Identify the
Functional Groups

14. The Experimental Design
About 5 min.

15. Experimental Design: Tissue Specificity
Class Neuron
Class Gland
Class Bone
Class Blood
Affy ID
Neuron
Gland
Bone Osteoblast
Blood White Cell
98063_at
1138.4
127.1
54.3
26.0
100080_at
17.0
592.5
27.2
372.8
103012_at
672.4
792.9
510.9
850.3 …
Expression Matrix
Expression Signal

16. Experimental Design: Disease
Disease state
Wild Type
Heart Disease
Time
08 w
16 w
24 w
Heart Disease 3
Wild Type
Experimental Design Matrix
Number of replicates

17. Important Points Clearly define your biological question(s)
Replicate experiments
Biological variability must be factored-in through replication:
repeat the experiment using different biological samples
Use clear and balanced designs
Use the same number of replicates in every class
Minimize experimental variability
Experimental variability arises from different platforms, different protocols, different experimenters, different days, etc…
Minimize all these factors
Control the assumptions of your design
Many studies on human patients assume two-class designs; however, the patients may exhibit heterogeneous phenotypes (e.g. different cancer stages) and hence different transcriptomes
The Explorative Analysis might reveal a different picture than you expected

18. Generating the Expression Signals
About 5 min.

19. Oligonucleotide Microarray

20. Oligonucleotide Microarray Technology
A transcript is recognized by probe pairs 25 nucleotides long
Raw fluorescence image

21. Oligonucleotide Microarray: Primary Signals
After essential image processing, we have signals for every probe
We need to integrate those signals into transcript/gene signals
Different techniques are available; two of the most popular ones are:
(MAS-5) detection p-value
p-value on a test of presence/absence of the transcript
relies on perfect match (PM) and mismatch (MM) probe signals
used for tissue-specificity or for pre-filtering
rma signal
continuous signal
relies only on perfect match probe signal
pre-normalized (no further normalization required)
used for differential expression (i.e. comparing two or more classes)

22. Enter the Matrix…

23. The explorative Analysis
About 15 min.

24. Aims of the Explorative Analysis
Quality Control
Are the samples directly comparable…
or are they affected by systematic biases?
 explore the signal distributions of the samples
Experimental Design (and beyond)
are the samples grouped according to the classes entailed by the experimental design?
are replicated experiments similar enough?
 use dimensionality reduction techniques (e.g. clustering, PCA, MDS) to explore global patterns

25. Explore the Distributions
Distributions can be explored using boxplots
Boxplots enable to visually compare many distributions at once
outliers
max point satisfying:
(x - Q3) < 1.5 * IQR
3rd quartile
median
1st quartile

26. Explore the Distributions
An example with real data

27. Explore the Distributions
What should we do if we see differences in the distributions?
Moderate differences can be corrected by normalization
(addressed in the pre-processing section)
Very large differences may be a sign of quality problems
use other diagnostics (e.g. look at the raw image files)
repeat single experiments
discard certain samples

28. Hierarchical Clustering
Hierarchical clustering enables to summarize the (dis)similarity structure of the samples
which samples are most similar and can be grouped together
what are the similarity relations between such groups
The distance is proportional to dissimilarity

29. Hierarchical Clustering
Hierarchical clustering can reveal sample anomalies
Somebody had fun the night before the experiment…

30. Hierarchical Clustering
Hierarchical clustering can reveal poor separation between classes

31. Hierarchical Clustering Technical Notes
Choose accurately the dissimilarity score
1 - Pearson Correlation
Euclidean
Make sure you have normalized samples

32. Heat Map
It is common to use a heat-map in combination with hierarchical clustering
Due to visualization limit, it is common not to use all the genes, but only the most differential ones
Caveat: the most differential genes may have a sample-clustering pattern different than the global one
Heat-maps can also be used to display the patterns of gene-sets

33. Principal Component Analysis (PCA)
PCA projects the data into a new data-space
the new dimensions are ranked by the amount of variance “explained”
the top-ranked dimensions can be picked for visual exploration

34. Dimensionality Reduction by Projection
The objects in a 3D space
Reduction to 2D space

35. Principal Component Analysis (PCA)
In microarray explorative analysis
Samples are treated as objects
Genes/transcripts are treated as dimensions
samples
samples
Principal Components
genes

36. Principal Component Analysis (PCA)
It is common to visualize the first two components in a bi-plot
unfortunately, the number of components that can be visualized altogether is limited
empirical approaches can be used to evaluate the number of “informative” principal components

37. Hierarchical Clustering vs PCA
Hierarchical clustering (HC) of samples and PCA can display partially different pictures
Cons of HC
More sensitive to noise
Assumes binary aggregations
Not suitable for time-course designs
Cons of PCA
Only 2-3 dimensions can be displayed simultaneously

38. Pre-processing
About 10 min.

39. Sample Normalization
Sample normalization can be used to correct global biases in gene signals
“Sample” because after normalization sample distributions will look more similar
Normalization, like all data transformations, must be used thoughtfully
Different levels
Equalization of the Central Value (Mean or Median)
Equalization of the Central Value and Spread (Standard Dev. or IQR)
Equalization of the Distribution Shape
Quantile Normalization

40. Sample Normalization Equalization of the Central Value
Equalization of the Central Value and Spread
Note: the mean (μ) can be replaced by the median, the standard deviation (sd) can be replaced by the Inter-Quartile Range (IQR)

41. Quantile NormalizationG A 97 72 50 B A F 81 45 41 E G A 97 72 50 B A F 97 72 50
1. Sort the distributions
2. Replace values

42. Quantile Normalization
After quantile normalization, the distributions look exactly the same

43. A Real-world Example Which is the normalized data-set?
What normalization did I use?
Are the distributions identical after normalization?

44. Gene Signal Standardization
Standardization can be used to make the gene signal scales (i.e. ranges) comparable
It is a transformation commonly used:
Before PCA (often done automatically by the software routine)
Before gene clustering
Before mapping to the heatmap color-scale

45. Differential Gene Expression
About 15 min.

46. Differential Gene Expression
The majority of experimental designs are two-class comparisons,
or can be broken down to two-class comparisons
E.g. treated vs. untreated, transgenic vs. wild-type
For such designs it is interesting to identify genes displaying different signals in the two classes (differential gene expression)

47. Differential Expression Scores
Oriented to Pure Strength:
Absolute Difference
Ratio of classes
Fold-change
Oriented to Statistical Significance:
t-test
Signal-to-noise
SAM (Significance Analysis) -- recommended
These scores can be used to:
Select gene-sets
Prioritize gene lists
Input for the identification of differential functional groups

48. Differential Expression Statistics
Oriented to Pure Strength
these statistics focus only on the magnitude of the change, but not on its consistency across replicated experiments
Oriented to Statistical Significance
these scores take into account the consistency of change across replicates; genes/transcripts with small but consistent changes can receive relatively high scores; however, they are usually preferable

49. Differential Expression Statistics
Absolute Difference
t statistic
Signal-to-noise
SAM statistic
Stabilizing constant
[SAM] Tusher VG, Tibshirani R, Chu G.
Significance analysis of microarrays applied to the ionizing radiation response.
Proc Natl Acad Sci U S A Apr 24;98(9): (PMID: )

50. From Statistics to Statistical Significance
Statistical significance is often expressed in the form of a p-value
When we compute a statistic (e.g. t statistic, SAM statistic) we then have to compute a p-value
p-values can be compared across different experiments, and
p-value thresholds can be directly related to false positive incidence
For the t-test, we use the a-priori know distribution of the t-statistic
For the SAM statistic, we have to use a permutation approach

51. SAM Permutation Approach
Class A
Class B
Class A
Class B
Permuted (rand)
Real

52. Functional Groups
About 30 min.

53. Identify the Functional Groups
Different Strategies
GENE SETS
PATHWAYS
NETWORKS
Spindle
Gene.1
Gene.2
Gene.3
P53 signaling
Gene.2
Gene.4
Gene.5
Score the set depending on the gene expression of its member genes
Just visual, or
Score the pathways exploiting gene expression and topology
Identify sub-networks (i.e. modules) satisfying some joint gene expression and topology requirement

54. Gene-set Enrichment: Competitive vs. Self-contained
Two different strategies for enrichment:
Self-contained
A differentiality statistic is computed for the gene-set
The statistical significance is evaluated by shuffling the columns of the gene expression matrix, and re-computing the differentiality statistic
Competitive
The enrichment of the gene-set is evaluated in comparison to the entire data-set, or random samples of genes (of the same size)

55. Testing Gene-sets: Fisher’s Exact Test / Hypergeometric Test
Two-Class
or Clusters
Is the intersection larger than expected by random sampling? UP
Threshold-dependent!!
Gene-set Collection

56. Testing Gene-sets: GSEA (Gene-Set Enrichment Analysis)
Statistics based on the cumulative sum-of-ranks
ESSet = Max (ES)
Weighting options
P-value and FDR estimated using permutations
Randomly sample gene sets
Shuffled phenotype labels

57. Competitive vs. Self-contained
How would you consider the Fisher’s Test and GSEA?

58. Testing Gene-sets: GSEA (Gene-Set Enrichment Analysis)
Statistics based on the cumulative sum-of-ranks
ESSet = Max (ES)
Weighting options
P-value and FDR estimated using permutations
Randomly sample gene sets (competitive)
Shuffled phenotype labels (hybrid)

59. Enrichment Maps
About 10 min.

60. GO.id GO.name p.value covercover.rat Deg.mdn Deg.iqr
GO: taxis E
GO: chemotaxis E
GO: adaptive immune response based on somatic recombination E
GO: adaptive immune response E
GO: leukocyte mediated immunity
GO: B cell mediated immunity
GO: myeloid cell differentiation
GO: immune effector process
GO: regulation of phagocytosis
GO: positive regulation of phagocytosis
GO: lymphocyte mediated immunity
GO: growth factor binding
GO: protein polymerization
GO: endoplasmic reticulum membrane
GO: immunoglobulin mediated immune response
GO: heart development
GO: response to bacterium
GO: regulation of endocytosis
GO: acute inflammatory response
GO: positive regulation of endocytosis
GO: myeloid leukocyte activation
GO: amino acid biosynthetic process
GO: regulation of inflammatory response
GO: activation of immune response
GO: positive regulation of immune system process
GO: positive regulation of immune response
GO: antigen processing and presentation
GO: regulation of immune system process
GO: regulation of immune response
GO: negative regulation of enzyme activity
GO: phagocytosis
GO: myeloid leukocyte differentiation
GO: humoral immune response
GO: lymphocyte activation
GO: leukocyte chemotaxis
GO: negative regulation of protein kinase activity
GO: negative regulation of transferase activity
GO: transforming growth factor beta receptor signaling pathw
GO: insulin-like growth factor binding
GO: T cell activation
GO: humoral immune response mediated by circulating immunogl
GO: cytosolic ribosome (sensu Eukaryota)
GO: protein amino acid N-linked glycosylation
GO: positive regulation of multicellular organismal process
GO: chemokine receptor binding
GO: chemokine activity
GO: Wnt receptor signaling pathway

61. Re-organizing the Gene Ontology
Gene Ontology is hierarchical, and terms are highly redundant / inter-related / inter-dependent
Enrichment Maps are not hierarchical, yet they neatly group redundant / inter-related / inter-dependent terms

62. Gene-set Overlap Measures
Jaccard Coefficient
Overlap Coefficient

63. Enrichment Map Visual StyleUP
Correlation to HD phenotype
DOWN
Anti-correlation to HD phenotype

65. Immune Cell Proliferation AcCoA Metabolism / Krebs Cycle
Cell Differentiation
Immune Cell Proliferation
AcCoA Metabolism / Krebs Cycle
Carbohydrate Metabolism / Glycosylation
Endomembrane System
Immune Response
Aminoacid Metabolism
NFkB
Phagocytosis
Coagulation
Oxidative Metabolism / Mitochondrion
Fatty Acid Metabolism
Peroxisome
Cell Motility
Antigen Recognition
Vacuole / Lysosome
Mitochondrial Ribosome
Metabolism
Heart Contraction / Blood Pressure Regulation
Protein Folding
Adherens Junctions
Ubq-dependent Protein Degradation
Growth Factor
Extracelluar Matrix
Embryonic Development
Apoptosis
Adhesion / Matrix /
Tissue Remodeling
RNA Processing / Translation
Bone / Cartilage Development
Protease Inhibitor
Angiogenesis
Tyr Kinase / Phosphatase
Phospho-inositide
Ruffle
Actin Cytoskeleton Remodeling
Cytoskeleton / Cell Cycle
Miscellanea
Microtubule Cytoskeleton
Mitotic Cell Cycle
Ras/Rho Signaling

66. Further Reading Allison DB, Cui X, Page GP, Sabripour M.
Microarray data analysis: from disarray to consolidation and consensus.
Nat Rev Genet Jan;7(1): Review.
PMID:
D'haeseleer P.
How does gene expression clustering work?
Nat Biotechnol Dec;23(12):
PMID:
Nam D, Kim SY.
Gene-set approach for expression pattern analysis.
Brief Bioinform May;9(3):
PMID:
Dinu I, Potter JD, Mueller T, Liu Q, Adewale AJ, Jhangri GS, Einecke G, Famulski KS, Halloran P, Yasui Y.
Gene-set analysis and reduction.
Brief Bioinform Jan;10(1):24-34.
PMID:
Cline MS, Smoot M, Cerami E, Kuchinsky A, Landys N, et al. (2007)
Integration of biological networks and gene expression data using Cytoscape.
Nat Protoc 2:
PMID:

67. Contact and Links Email daniele.merico@gmail.com Web-site