Introduction to Microarray Analysis

Download Report

Transcript Introduction to Microarray Analysis

Introduction to
Microarray Analysis
Uma Chandran PhD, MSIS
Department of Biomedical Informatics
[email protected]
412-648-9326
10/17/12
What is a microarray

Probes on surface


Arrays can detect





Glass beads, chips, slides
mRNA
microRNA
Methylation
SNP
High throughput


10000s of specific probes
Measure global gene
expression, SNP calls,
LOH, amplification,
methylation etc
Questions that can be asked




Can measure global changes
Which mRNAs are high in disease versus
normal, i.e, out of the 1000s of mRNAs
expressed in the cell at any time
Are there single nucleotide polymorphism
that are markers for a disease – many
studies on for example, autism,
schizophrenia
Are there methylation changes in disease
versus normal
ARRAY DESIGN
Affymetrix
Insert oligo slide

Probes are synthesized
on a chip

Probes are oligonculeotides
of a specified length


Generally 25 mers
At each x, y location a
particular oligonucleotide
is synthesized in 1000s of
copies at that location
Affymetrix
•
•
•
Feature: a location on the array with a particular oligonucleotide sequence
Oligonucleotides are synthesized using a photolithographic manufacturing
• process
The oligo on the chip is called the probe and RNA (or DNA) that it hybridizes to
is called the target
Affy array design
Probe set
Affymetrix
Probe design


Multiple probe sets/gene
Probe sets are selected based on





GenBank
dbEST
RefSeq
Bioinformatics approaches
Design at the time of chip design

However, this may be incorrect as genome builds
update
Affymetrix data
Annotation




The probe set id and
sequence are contained
in reference files
This id never changes
However, annotations
change with genome
builds
Many software tools to
annotate


Some involve new BLAST
of the sequences
Mask out probe sets
Affymetrix

Chips for

Human

HGU95, HGU133A, B, HGU133 set










Very low ~ 10 units
20K +
Cannot compare genes within chips




Mouse
Rat
Chimpanzee
Plants
Many other species
Dynamic range


54K probe sets on the HGU133, 30+ to known genes and ESTs
Control probes like GAPDH
Spike in bacterial probes
For example, a transcript that is expressed at 500 units may not be more abundant than one
that is expressed at 200 units
This is due to probe binding affinities etc
However, can compare the same probe across multiple chips
Difficulty in probe design makes it difficult to compare from one version to another
Affymetrix workflow
from: http://wwwnmr.cabm.rutgers.edu/academics/biochem694/reading/DalmaWeiszhau
sz_2006.pdf
Illumina
Illumina
Each bead has one type
of oligo and thousands of
these oligos/bead
Bead is deposited on
wells in glass slides. The
beads are decoded by a
step by proprietary
technology
Microarray analysis objectives
Data Preprocessing
Data Analysis
Analysis questions
Treatment

Class Comparison



Class Discovery



Expression - Which genes/miRs
are up or down in tumors v normal,
untreated v treated
SNP – Which regions are amplified
or deleted
Within the tumor samples, are
there subgroups that have a
specific expression profile?
SNP – amplification or deletion
common to subgroups?
Class prediction, pathway analysis
etc

Integrative analysis



Proteomic and genomic
SNP and expression
Methylation and expression
Normal
Challenges in microarray
analysis

Different platforms
 Ilumina, Affymetrix, Agilent….



Many file types, many data formats
Need to learn platform dependent methods and software required
Analysis
 How to get started?
 Which methods? Which software? Many freely available tools.
Some commercial
 Analysis software and methods will depend on platform.




SNP analysis is different from expression
Software used may be very specific to SNP
For example, Excel cannot open large SNP files
How to interpret results
Public databases




Many sources for public
data – labs, consortia,
government
Publications require that
data files including raw files
be made public
GEO –
http://www.ncbi.nlm.nih.gov/
geo/
Array Express http://www.ebi.ac.uk/arraye
xpress/#ae-main[0]
Hands on #1





Look at GEO
Search Data Set with the term Exercise
Exercise Heart Human
Identify Platform by clicking on GSE record
Try restricting by platform such as Affymetrix
or Illumina
Affy data
Normalization method
Signal value
Probe set Id
Total probesets
Raw files
Data pre-processing



Affy produces many files - .dat, .cel, .chp etc
Process these to produce data that can be
opened in excel or .txt
Illumina produces different file types
Data Preprocessing

Objective


Convert image of
thousands of signals to a a
signal value for each gene
or probe set
Multiple step

Image analysis



Background and noise
subtraction
Normalization
Summarized expression
value for a probe set or
gene
Gene 1
Gene 2
Gene 3
.
Gene10000
100
150
75
500
Data Pre-processing






Go from .DAT file to feature
quantification
The first step where .DAT file
is aligned to a grid and the
features are quantified is
usually performed by Affy’s
proprietary algorithm
.DAT
.CEL file
.CEL file contains the feature
quantifications
.CEL file still has probes
spread over the chip
Values still need to be
summarized to probe set level;
for example 90525_at = 250
units
250
Data Pre-processing – Step 1

Image processing


Usually done using proprietary software
Affy: convert .dat file to .cel file


May perform noise subtraction, background
Illumina: Bead Studio software to convert bead
level data to next level of data
Data Preprocessing – Step 2

Normalization





Bring all the experiments
up to the same scale
Multi-step process
depending on technology
Summarized expression
value for a probe set or
gene
Affy: .cel to .chp; need .cdf
file which describes the file
layout
Ilumina: normalization
option and background
subtraction option using
Bead Studio
Gene 1
Gene 2
Gene 3
.
Gene10000
100
150
75
500
.CEL +.CDF to .CHP

In going from .CEL to
.CHP file to generate
signal values, the
multiple probes within
a probe set are
“averaged” to produce
a single value for that
gene/transcript
Normalization




Corrects for variation in
hybridization etc
Important for all high
throughput platforms
Assumption that no global
change in gene expression
Without normalization


Treated
Intensity value for gene will
Gene 1
100
be lower on Chip B
Gene 2
150
Many genes will appear to Gene 3
75
be downregulated when in .
reality they are not
Gene10000 500
Control
50
75
32
250
How to normalize?

Many methods – Affy MAS5.0

Median scaling – median intensity
for all chips should be the same

Known genes, house keeping,
invariant genes
Quantile - RMA


Normalization method may
differ depending on platform


Illumina – cubic spline
Affymetrix



Which method to choose?


Choose method
.cel to .chp file
Know the biology
After normalization from
.cel


.chp file
.txt file
A
Before
After
100
200
B
50 (down)
200 (no change)
Normalization
Affy data
Normalization method
Signal value
Probe set Id
Total probesets
Raw files
Workflows

normalization
Affy
 .dat file > .cel file > .chp file > .txt file
cdf file
Affy software needed for .dat > cel
 The rest of the steps can be carried out by other tools
Illumina
 Through Bead Studio




Bkg subtraction > normalization with various options > background
normalization > .txt file
Need bead studio to carry out these steps and raw files not
necessarily given
Illumina



Does not have .DAT, .CEL, .CDF and .CHP
files
There is no chip definition or chip layout as in
Affy
However, the identity of each bead has to be
decoded vial proprietary software
Illumina
Data preprocessing
Signal
normalization
Raw files are .txt files
Probe id
Affy v Illumina

Affy








25mer
Probe synthesized on chips
Multiple probes/probeset
May have multiple
probes/transcript
.dat, .cel, .cdf, .chp file
types
Normalization methods
such as quantile
Txt output can be used for
downstream data analysis
Annotations can be updated

Illumina








Longer oligo
Bead technology
Single probe
May have multiple
probes/transcript
Image file processed by
Bead Studio
Several normalization
methods
Txt output can be used for
downstream data analysis
Annotations can be updated
Hands on #2 -Data analysis


Import data into BRB
Which files to import


.cel file if performing normalization through BRB
Or mport already normalized file as .txt file for
further analysis
Steps in analysis - Import

Affy





Import all files into Affy tools such as Expression console
Normalize and generate signal values using Affy MAS5.0
Assess QC using GAPDH, B-actin and control probes for
spike in and hybridization
Then, import into other tools such as BRB for analysis
Illumina




Depending on background subtraction/normalization, may
have generated negative values
Check QC metrics, such as did chip pass?
Remove negative values
Import into tools such as BRB
Step in Data analysis –
Normalization


Import raw data into a tool
Has data been normalized?



After normalization, check distribution



If not, which method to use? What is available for a
particular platform
If not available in tools, is R code or package available
Are there any batch effects?
Is the data log transformed?
 If not, should you log transform? When? After or before
normalization?
Are there missing or negative values in data?

What should be done? Impute? Remove rows
Steps in Data analysis –
update Annotations




Very important step
Annotations updated
Annotations provided
may often be incorrect
Multiple probe sets for
each gene
BRB – Array tools



Website
Excel plug in; R and fortran
Import, choose correct format


For Affy:
.cel files


Or directly from processed files


Process using GCRMA or MAS5.0
Attaches annotation
Create experiment labels
Class Discovery

Objective?




Can data tell us which classes are similar?
Are there subgroups?
Do T-ALL, T-LL, B-ALL fall into distinct groups?
Methods



Hierarchical clustering
K-means, SOM etc
These are Unsupervised Methods
 Class Ids are not known to the algorithm


For example, does not know which one is cancer or non cancer
Do the expression values differentiate, does it discover new
classes
Multidimensional scaling MDS
Class comparison – differential
expression analysis


What genes are up
regulated between control
and test or multiple test
conditions
 Normal v tumor
 Treated v untreated
Fold change


Not sufficient, need
statistics
Statistics

t test, non-parametric, fdr,
Class comparison

Many analysis methods
 May produce different results
 Different underlying statistics and methods







t test
t test with permutations
SAM
Emperical bayesian
Depends on underlying assumptions about data
High throughput data with many rows and few samples
 What is the distribution
 Variance from gene to gene
Save raw data files to try different methods and compare results
Fold change does not take variation
into account
low
variability
medium
variability
high
variability
Modified from madB
http://nciarray.nci.nih.gov/
Differentially expressed gene
Differentially expressed gene.
A low-reliable estimate
Differentially expressed
gene. Powerful and exact
statistical tests must be used
Hypothesis Testing
Normal
Tumor
d
mean1 mean2
Null hypothesis
Alternative hypotheses
Statistical power

t test


Test hypothesis that the two
means are not statistically
different
Adding “confidence” to the fold
change value
 Mean
 Standard deviation
 Sample size
 Calculates statistic
 You choose cutoff or
threshold

Give me gene list at a cutoff of p
<0.05

95% confidence that the
mean for that gene between
control are treated are
different
Experimental Design – Very
important!!!

Sample size

How many samples in test and
control



Will depend on many
factors such as whether
tissue culture or tissue
sample
Power analysis
Replicates


Technical v biological
 Biological replicates is more
important for more
heterogenous samples
Need replicates for
statistical analysis
To pool or not to pool


Sample acquistion or
extraction


Depends on objective
Laser captered or gross
dissected
All experimental steps from
sample acquisition to
hybridization

Microarray experiments are
very expensive. So, plan
experiments carefully
t tests

Results might look like


At a p<0.05, there are
300 genes up and 200
genes downregulated
 95% confidence that the
means of these genes
in the two groups is
different
At a p < 0.05, x genes up
and y genes down with a
fold change of at least
3.0
Multiple comparison



Microarrays have multiple comparison problem
p <= 0.05 says that 95% confidence means are
different; therefore 5% due to chance
5% of 10000 is 500







500 genes are picked up by chance
Suppose t tests selects 1000 genes at a p of 0.05
500/1000 ;Approximately 50% of the genes will be false
Very high false discovery rate; need more confidence
How to correct?
Correction for multiple comparison
p value and a corrected p value
Corrections for multiple
comparisons




Involve corrections to the p value so that the
actual p value is higher
Bonferroni
Benjamin-Hochberg
Significance Analysis of Microarrays

Tusher et al. at Stanford
Hands on BRB

Class comparison
 Choose comparison
 Which tests are available?
 P value cutoff
 How is multiple correction
testing being done?





Stringent p value, fdr
How is the output reported?
Can you figure out how many
genes are regulated at
different p values and different
cutoffs
How to interpret results
Look at gene lists generated
by our analysis v those
generated in the paper
BRB – Class Comparison








Output folder
Check the .html file
Look at results
P value
Fold change
Annotation
Click on annotation
Cut and paste save into Excel
Issues

Annotation








Multiple probe sets for a gene
Annotation files will get updated
Which one is correct?
Where does it map?
How to report the genes?
How to compare between platforms
Different chips within same platform
Biological annotation
Difficult to interpret
experimental results
350
4500
201120_s_at progesterone
receptor membrane component 1
PGRMC1
4000
300
204253_s_at vitamin D (1,25dihydroxyvitamin D3) receptor
VDR
250
200
204254_s_at vitamin D (1,25dihydroxyvitamin D3) receptor
VDR
150
204255_s_at vitamin D (1,25dihydroxyvitamin D3) receptor
VDR
213692_s_at Vitamin D (1,25dihydroxyvitamin D3) receptor
VDR
100
50
201121_s_at progesterone
receptor membrane component 1
PGRMC1
3500
3000
201701_s_at progesterone
receptor membrane component 2
PGRMC2
2500
208305_at progesterone receptor
PGR
2000
1500
213227_at progesterone receptor
membrane component 2 PGRMC2
1000
228554_at progesterone receptor
PGR
500
0
0
1
10 19 28
37 46 55 64
73 82 91 100 109 118 127 136
1
10 19 28 37 46 55 64 73 82 91 100 109 118 127 136
200
100
0
Unlogged Expression value
300
Which probe/probe set is
correctly aligned to the gene?
205225_at
211233_x_at
211234_x_at
211235_s_at
211627_x_at
Affymetrix probeset
215551_at
215552_s_at
217163_at
217190_x_at
Probe set errors
Types of Probe Error
Cross
Hybridization
Mismatched
Probe
Intron Probe
SNPs
ESR1 probes in UCSC
genome browser
How to manipulate Gene lists

Create gene lists



Venn Diagram
Can be done even though study done on different
platforms
Compare MAS and RMA


Venn Diagram
Compare B-ALL v T-LL and T-LL v B-ALL
Venn Diagram
http://www.pangloss.com/seidel/Protocols/venn.cgi
http://ncrr.pnl.gov/software/VennDiagramPlotter.stm
Conclusion

Other analysis





Class prediction
Gene list from class comparison can be used in
pathway analysis
HSLS pathway workshops on Ingenuity, DAVID,
Pathway Architect
Future:
 Integrate expression data with other data such as
snp or microRNA
GEO has some data analysis features
ESR1 probes in UCSC
genome browser
Next Gen Sequencing

Directly sequence DNA to determine







SNP
CN
Expression, mRNA, microRNA
Protein binding sites
Methylation
Initial steps depend not on hybridization but also
on base pairing or complementarity and DNA
synthesis
Data analysis extremely challenging
Next Gen Sequencing
Applications






Sequence varation – WGS, Exome Seq
Structural rearrangements – WGS, Exome
Seq
Copy number – WGS, Exome Seq
Epigenetic changes such as methylation –
Methyl Seq
DNA – protein binding – CHIP Seq
mRNA expression – RNA Seq
Next Gen Sequencing
Read mapping
Alignment


Denovo assembly
Mapping to reference
genome


Based on complementarity
of a given 35 nucleotide to
the entire genome
Computationally intensive


Million of 35 bp reads has to
search for alignment against
the reference and align
spefically to a given regions
Large file sizes


Sequence files in the TB
Aligned file BAM files

Several hundred GB
Reference genome
Sequence variation
Analysis pipeline- CHIP-Seq