1. Previous Lecture: Proteomics Informatics
Example data – MALDI-TOF
Peptide intensity vs m/z

2. Gene Expression Analysis (I)
This Lecture
Gene Expression Analysis (I)

3. Learning Objectives Microarray experimental details
Microarray data formats
QC analysis and data exploration
Normalization
Differential expression
Functional enrichment
Databases

4. The Central Dogma of Molecular Biology DNA is transcribed into RNA which is then translated into protein
protein
RNA
DNA
transcription
translation
replication
Measured by Microarray

5. What is a Microarray A simple concept: Dot Blot + Northern
Reverse the hybridization - put the probes on the filter and label the bulk RNA
Make probes for lots of genes - a massively parallel experiment
Make it tiny so you don’t need so much RNA from your experimental cells.
Make quantitative measurements

6. Microarrays are Popular
At NYU Med Center we are now collecting about 3 GB of microarray data per week (60 chips, 6-10 different experiments)
PubMed search "microarray"= 13,948 papers
2005 = 4406
2004 = 3509
2003 = 2421
2002 = 1557
2001 = 834
2000 = 294

7. A Filter Array

8. DNA Chip Microarrays
Put a large number (~100K) of cDNA sequences or synthetic DNA oligomers onto a glass slide (or other subtrate) in known locations on a grid.
Label an RNA sample and hybridize
Measure amounts of RNA bound to each square in the grid
Make comparisons
Cancerous vs. normal tissue
Treated vs. untreated
Time course
Many applications in both basic and clinical research

9. cDNA Microarray Technologies
Spot cloned cDNAs onto a glass microscope slide
usually PCR amplified segments of plasmids
Label 2 RNA samples with 2 different colors of flourescent dye - control vs. experimental
Mix two labeled RNAs and hybridize to the chip
Make two scans - one for each color
Combine the images to calculate ratios of amounts of each RNA that bind to each spot

10. Spot your own Chip (plans available for free from Pat Brown’s website)
Robot spotter
Ordinary glass microscope slide

12. Combine scans for Red & Green
False color image is made from digitized fluorescence data, not by superimposing scanned images

13. cDNA Spotted Microarrays

15. Data Acquisition Scan the arrays Quantitate each spot
Subtract background
Normalize
Export a table of fluorescent intensities for each gene in the array

16. Affymetrix “Gene chip” system
Uses 25 base oligos synthesized in place on a chip (20 pairs of oligos for each gene)
RNA labeled and scanned in a single “color”
one sample per chip
Can have as many as 20,000 genes on a chip
Arrays get smaller every year (more genes)
Chips are expensive
Proprietary system: “black box” software, can only use their chips

17. Affymetrix Gene Chip

20. Affymetrix Technology

23. Affymetrix Software Affymetrix System is totally automated
Computes a single value for each gene from 40 probes - (using surprisingly kludgy math)
Highly reproducible (re-scan of same chip or hyb. of duplicate chips with same labeled sample gives very similar results)
Incorporates false results due to image artefacts
dust, bubbles
pixel spillover from bright spot to neighboring dark spots

24. Affymetrix Pivot Table

25. Plot of raw data (PM probes)

26. Plot of log2 data (PM probes)

27. MA plot: log of fold change (M) vs log of Intensity (A)
M = log2 (A/B)
A = ½ log2 (A*B) = ½ (log2 (A) + log2 (B))
Hypox1 vs
Hypox2
Hypox3
Norm1
Norm2
Norm3

28. Goals of a Microarray Experiment
Find the genes that change expression between experimental and control samples
Classify samples based on a gene expression profile
Find patterns: Groups of biologically related genes that change expression together across samples/treatments

29. Basic Data Analysis
Fold change (relative increase or decrease in intensity for each gene)
Set cutoff filter for low values (background +noise)
Cluster genes by similar changes - only really meaningful across multiple treatments or time points
Cluster samples by similar gene expression profiles

30. Streamlined Affy Analysis
Normalize
Filter
Raw data
•Present/Absent •Minimum value •Fold change
(RMA)
Significance
Classification
Clustering
•t-test
•SAM
•Rank Product
•PAM
•Machine learning
Gene lists
Function
(Genome Ontology)

31. Scatter plot of all genes in a simple comparison of two control (A) and two treatments (B: high vs. low glucose) showing changes in expression greater than 2.2 and 3 fold.

32. Thomas Hudson, Montreal Genome Center

33. Normalization
Can control for many of the experimental sources of variability (systematic, not random or gene specific)
Bring each image to the same average brightness
Can use simple math or fancy -
divide by the mean (whole chip or by sectors)
LOESS (locally weighted regression)
No sure biological standards

34. RMA Robust Multichip Average log(medpol(PMij − BG)) = µ i + α j + e ij
Bolstad, B.M., Irizarry R. A., Astrand, M., and Speed, T.P. (2003), A Comparison of Normalization Methods for High Density Oligonucleotide Array Data Based on Bias and Variance. Bioinformatics 19(2):
log(medpol(PMij − BG)) = µ i + α j + e ij
for (array i, probe j)

35. Are the Treatments Different?
Analysis of microarray data has tended to focus on making lists of genes that are up or down regulated between treatments
Before making these lists, ask the question: "Are the treatments different?"
PCA/MDS or cluster the samples
If the treatment is responsible for differences, then use statistical methods to find the genes most responsible
If there are not significant overall differences, then lists of genes with large fold changes may only reflect random variability.

36. Statistics
When you have variability in measurements, you need replication and statistics to find real differences
It’s not just the genes with 2 fold increase, but those with a significant p-value across replicates
Non-parametric (i.e. rank or permutation) or paired value statistics may be more appropriate (low number of samples, high standard deviation)

37. Multiple Comparisons
In a microarray experiment, each gene (each probe or probe set) is really a separate experiment
Yet if you treat each gene as an independent comparison, you will always find some with significant differences
(the tails of a normal distribution)
Different genes are NOT independent

38. False Discovery
Statisticians call false positives a "type 1 error" or a "False Discovery"
The FDR must be smaller than the number of real differences that you find - which in turn depends on the size of the differences and variability of the measured expression values
You can’t know the true false discovery rate for your data, but it can be estimated in a number of different ways.
In biology we tend to be comfortable with an estimated FDR of 5-10%

39. SAM Significance Analysis of Microarrays R package, Excel plugin Free
Tusher, Tibshirani and Chu (2001): Significance analysis of microarrays applied to the ionizing radiation response. PNAS : , (Apr 24).
R package, Excel plugin
Free
Permutation based
Most published method of microarray data analysis

40. SAM- procedure overview
Sample genes
expression
scale
Define and calculate
a statistic, d(i)
Generate permutated
samples
Estimate attributes
of d(i)’s distribution
Identify potentially
Significant genes
Choose Δ
Estimate FDR

41. Calculate “relative difference” – a value that incorporates the change in expression between conditions and the variation of measurements in each condition
Calculate “expected relative difference” – derived from controls generated by permutations of data
Plot against each other, set cutoff to identify deviating genes
Calculate FDR for chosen cutoff from the control permutations

42. Relative Difference Mean expression of gene i in conditions I and U
Gene-specific scatter
Constant to reduce variation of low expressed genes

43. SAM Two-Class Unpaired
Permutation tests
For each gene, compute the d-value (similar to a t-statistic). This is the observed d-value (di) for that gene.
ii) Randomly shuffle the expression values between groups A and B. Compute the d-value for each randomized set.
iii) Take the average of the randomized d-values for each gene. This is the ‘expected relative difference’ (dE) of that gene. Difference between (di) and (deE) is used to measure significance.
iv) Plot d(i) vs. dE(i)
v) Calculate FDR = average number of genes that exceed in the permuted data.
Exp 1
Exp 2
Exp 3
Exp 4
Exp 5
Exp 6
Gene 1
Group A
Group B
Original grouping
Exp 1
Exp 4
Exp 5
Exp 2
Exp 3
Exp 6
Gene 1
Group A
Group B
Randomized grouping

44. SAM Two-Class Unpaired
Significant positive genes
(i.e., mean expression of group B >
mean expression of group A)
SAM Two-Class Unpaired
“Observed d = expected d” line
Plot d(i) vs. dE(i)
For most of the genes:
The more a gene deviates from the “observed = expected” line, the more likely it is to be significant. Any gene beyond the first gene in the +ve or –ve direction on the x-axis (including the first gene), whose observed exceeds the expected by at least delta, is considered significant.
Significant negative genes
(i.e., mean expression of group A > mean expression of group B)

45. Higher Level Microarray data analysis
Clustering and pattern detection
Data mining and visualization
Controls and normalization of results
Statistical validatation
Linkage between gene expression data and gene sequence/function/metabolic pathways databases
Discovery of common sequences in co-regulated genes
Meta-studies using data from multiple experiments

46. Types of Clustering Herarchical Self Organizing Maps (SOM)
Link similar genes, build up to a tree of all
Self Organizing Maps (SOM)
Split all genes into similar sub-groups
Finds its own groups (machine learning)
Principle Component
every gene is a dimension (vector), find a single dimension that best represents the differences in the data

47. Cluster by fold change

48. GeneSpring

50. SOM Clusters

51. Classification
How to sort samples into two classes based on gene expression data
Cancer vs. normal
Cancer sub-types (benign vs. malignant)
Responds well to drug vs. poor response (i.e. tamoxifen for breast cancer)

52. PAM: Prediction Analysis for Microarrays
Class Prediction and Survival Analysis for Genomic Expression Data Mining
Performs sample classification from gene expression data,
via "nearest shrunken centroid method'' of Tibshirani, Hastie, Narasimhan and Chu (2002):
"Diagnosis of multiple cancer types by shrunken centroids of gene expression"
PNAS : (May 14).

53. BioConductor
All of these normalization, statistical, and clustering methods are available in a free software package called BioConductor, which is part of the R statistical environment
command line interface
> data(SpikeIn)
> pms <- pm(SpikeIn)
> mms <- mm(SpikeIn)
> par(mfrow = c(1, 2))
> concentrations <- matrix(as.numeric(sampleNames(SpikeIn)), 20,
+ 12, byrow = TRUE)
> matplot(concentrations, pms, log = "xy", main = "PM", ylim = c(30, ))
> lines(concentrations[1, ], apply(pms, 2, mean), lwd = 3)
> matplot(concentrations, mms, log = "xy", main = "MM", ylim = c(30,
> lines(concentrations[1, ], apply(mms, 2, mean), lwd = 3)

54. Functional Genomics
Take a list of "interesting" genes and find their biological relationships
Gene lists may come from significance/classfication analysis of microarrays, proteomics, or other high-throughput methods
Requires a reference set of "biological knowledge"

55. Genome Ontology How to organize biological knowledge?
Biologists work on a variety of different research organisms: yeast, fruit fly, mouse, … human
the same gene can have very different functions (antennapedia)
and very different names (sonic hedgehog…)

56. GO
Biologists got together and developed a sensible system called Genome Ontology (GO)
3 hierarchical sets of terminology
Biological Process
Cellular Component (location within cell)
Molecular Function
about 1000 categories of functions

58. List (and convert) gene identifiers from many genomic resources including NCBI, PIR and Uniprot/SwissProt as well as Illumina and Affymetrix gene IDs
Gene IDs matched to GO function annotations (for human)
Test for enrichment of GO categories (or KEGG pathways, disease associations, etc.) in list.
Groups significant categories into clusters

59. DAVID enrichment score: EASE
DAVID uses a modified Fishers Exact text to get p-values for enrichment.
Basic idea: is enrichment of this category in this list greater than frequency of the category in the genome.
A Hypothetical Example: In human genome background (20,000 gene total), 40 genes are involved in p53 signaling pathway. A given gene list has found that 3 out of 300 belong to p53 signaling pathway. Then  we ask the question if 3/300 is more than random chance comparing to the human background of 40/20000.
Fisher Exact P-Value = 
However, EASE Score is more conservative. EASE Score =  0.06 (using 3-1 instead of 3). Since P-Value > 0.01, this user gene list is specifically associated (enriched) in p53 signaling pathway no more than random chance

60. Microarray Databases
Large experiments may have hundreds of individual array hybridizations
Core lab at an institution or multiple investigators using one machine - data archive and validate across experiments
Data-mining - look for similar patterns of gene expression across different experiments

61. Public Databases
Gene Expression data is an essential aspect of annotating the genome
Publication and data exchange for microarray experiments
Data mining/Meta-studies
Common data format - XML
MIAME (Minimal Information About a Microarray Experiment)

62. Array Express at EMBL

64. GEO at the NCBI

69. Sumary Microarray experimental details Microarray data formats
QC analysis and data exploration
Normalization
Differential expression
Functional enrichment
Databases

70. Next Lecture: Next Generation Sequencing Informatics