1
2
3
4	Microarray Technology Introduction Nucleic acid hybridization/antibody based methodology for expression or mapping studies high throughput screening method Utilizes very large collections of substrate immobilized DNA or Proteins Large scale and concurrent survey of Gene/Protein expression changes Gene expression/protein profiling Genomic Variation Single nucleotide polymorpism detection/ Haplotype analysis (HAPMap) requires bioinformatic approaches for data management and analysis
5	Expression Profiling Molecular description of a cell (system) at ultra high resolution Transcriptome Mapping Molecular Finger Printing Genetic Networks Transcriptional dependencies Inferential Pathway Analysis Signal transduction targets Systems biology tool
6	Some Terminology Probe = spotted DNA Target = labeled sample Feature = Probe Pitch = center to center spacing of probes Feature density = total number of probes Probes are not the same as genes Unigene ID = EST to EST association Entrez ID = sequence associated with genomic locus GO ID = functionally annotated genes
7
8
9	What do we need? Arrayer Substrates Membranes (³³P) Glass (Fluorescence) Beads (Fluorescence) Probe sets Oligonucleotides EST/cDNA Scanner RNA
10	Substrates and Arrays – what are the Options? Substrate linked Synthesis (Glass, beads) Affymetrix, Nimblegen (fluorescence) Light directed synthesis oligonucleotides Agilent (fluorescence) Spatially directed fluidics based synthesis oligonucleotides Illumina, Luminex (fluorescence) Bead linked synthesis oligonucleotides Contact Printing (glass, membranes) cDNA (fluorescence, ³³P) Oligonucleotides (fluorescence, ³³P) Proteins (antibodies, fluorescence)
11	Detection of Hybridization Events
12	Where do the we get the DNA to put on arrays? Oligos - designed based on sequence data ORFs - PCR primers designed based on sequence data, often use tailed primers cDNA - clones from various clone collections, PCR amplified and purified
13	Affymetrix (Nimblegen) GeneChip technology Short Oligonucleotides 25 mer (Affymetrix 50 mer (Nimblegen) 40-1300k DNA oligos on ~ 2.5 cm² glass surface. Expression arrays Human, Mouse, Rat, Yeast, E. coli, Drosophila, C. elegans, Dog, Soybean, Plasmodium/Anopheles, Pseudomonas, Arabidopsis, Zebrafish, Xenopus, etc. DNA analysis arrays Resequencing, SNP analysis, LOH Custom arrays
14	Genechip Synthesis The manufacturing of GeneChip^® probe arrays is a combination of photolithography and combinatorial chemistry. See: http://www.Affymetrix.com/Technology.html
15	Genechip Design
16	Agilent inkjet technology long (60mer) oligonucleotide arrays Arrays include Expression arrays (Human, Mouse, Rat, Arabidopsis, rice, Magnaporthe and yeast expression arrays) Promoter arrays (human and mouse) Custom arrays (Rapid turnaround,8.4K or 22K feature sizes, up to 8 fields per slide) Custom arryasAgilent also creates custom arrays
17	Chip Manufacturing Process
18	Multiplexed DNA Micro Array Analysis
19	Quantum Dot Properties
20	Decoding of Dye based Beads
21
22	Direct Hybridization RNA Profiling Bead 50 base gene-specific Probe linked to 23 base Address Hybridized to labeled nucleic acid made from total RNA Each bead coated with same probe oligo (100,000’s copies) ~30 copies of each bead type per array
23	Decoding of SEBs - Fiber Optics & Microwells
24
25
26
27
28
29
30
31	Sources of Variance Arrays Feature Consistency Spotting Volume Pin characteristics Substrate Homogeneity DNA binding capacity Environment Dust Humidity Aerosols Light Temperature Oxidants RNA Purity à Efficiency of cDNA synthesis Integrity à Length of FSR transcript Labeling
32
33	False-positive error and array experimental costs
34	Planning A Good Array Experiment Experimental Design What is the biological question, i.e. what comparisons should be made? What is the type of biological comparisons? How is sample complexity controlled? How many biological replicates are required? Data Analysis How will differentially expressed genes be identified? How will errors be estimated?
35	Available Solutions Most integrated and optimized: Commercial Software SAS, SPSS, S-Plus (general) Spotfire, GeneSight, GeneSpring (specific) Custom TM4, BAMarray, Powerarray, ClusFavor etc. (specific) Main Issues Proprietary Hidden data handling Most versatile, most recent and transparent data handling: Open Source BioConductor/R SAM SPH/EB-arrays LIMMA & Tcl GUI R/MANOVA & JAVA GUI D-Chip & C++ GUI Main Issues Data import File format Requires programmer’s support
36
37
38
39
40	Data Preprocessing How to remove systematical biases!
41
42	Background Correction None DNA vs Substrate No Imputation/Offset Local Negative Signal Intensities likely Imputation/Offset required Global Negative Signal Intensities likely Imputation/Offset required Moving Minimum 3x3 spot average background Negative Signal Intensities likely Imputation/Offset required
43
44
45	Two Color Arrays: Loess Normalization Intra Arrray Intensity dependent Model based Stat Tests (MAANOVA)
46	Two Color Arrays: Quantile Normalization Intra/Inter Array Factor based Removal of High Intensity Outliers Standard Stat Tests (SAM, SPH, LIMMA)
47	Beware Any data adjustment, be it performed as sophisticated or industrious as possible, cannot convert low quality data into high quality data Data adjustment always removes a part of the biology !!Use it as sparingly as possible!!
48	Statistical Analysis How to select differentially expressed Genes!
49	Identification of Differentially Expressed Genes Degree of Regulation Small changes Less reproducible Few genes @ Standard Significance threshold More replicates required Large Changes More reproducible Many genes @ Standard Significance threshold Fewer replicates required But: Effect size is gene dependent (Transcriptome Survey) Data preparation and Statistical Analysis
50
51	Concordance Analysis for Replication 11: ratios from A vs B comparison; replicate 1 12: ratios from A vs B comparison; replicate 2 13: ratios from B vs A comparison; replicate 3 15: ratios from B vs a comparison; replicate 4 Concordance coeff.: 0.947 – 0.961
52
53
54
55
56
57	Heuristic selection of genes I by median & standard deviation (CV)
58
59
60	Data Example
61
62
63	Why Doesn’t the t-statistics Work? When diff_g and are small, then is big (then p becomes very small). So gene g is more likely to be declared as a DE gene. This method is biased in favor of selecting genes with small diff_g and .
64
65	SAM Plot
66	Multi-Level Modeling:
67	Multi-Level Modeling T-tests (SAM) High accuracy data Minimal preprocessing artifacts Level1 Model (LIMMA, MAANOVA) Gene specific Y_obs =f(Y_true) Y_obs =g(Y_array) + g(Y_dye) + g(Y_batch) +………g(Y_treatment) Level2 Model (SPH, EBayes) Population specific Y_obs =f(Y_up) + f(Y_down) + f(Y_not) Borrow and share information appropriately for better estimates
68	Multi-Level Modeling: LIMMA, MAANOVA, SPH Ability to model various sources of variability: Level 1 model: Gene Specific Model Model the observed log intensities as a function of the unknown true log intensity. detailed modeling of experimental variability: within array, between array, estimation of gene specific variability … Level 2 model: Population Average Model All unknown quantities are given prior distributions Building of all these features into a common model Ability to borrow and share information in appropriate ways to get better estimates
69
70	"By computing the Analysis of..." By computing the Analysis of Variance (ANOVA), we can mathematically estimate the different sources of variation and systematically detect treatment effects in the data. The real question is: which genes are differentially expressed between the samples? In our framework, we ask which variety-by-gene (VG) effects are statistically significant. Thanks to the use of Model, you can omit certain preprocessing steps.
71
72	MAANOVA produces Four F-like Statistics F1g measures the gene specific treatment effect F3g measures the gth gene treatment effect using the pooled variance estimator F2g measures the gth gene treatment effect using both the gene specific variance estimator and the pooled variance estimator with equal weight. Fsg measures the gth gene treatment effect using a shrinkage variance estimator. Fs is the most robust and usually most powerful (Cui, Churchill et all)
73
74	Volcano Plot A ‘volcano’ plot provides a graphical summary of the simultaneous results from all four F-tests. On the plot, the y-axis value is -log10(P-value) for the F1 test. The x-axis value is proportional to the fold changes. A horizontal line represents the significance threshold of the F1 test. Blue dots: EE genes Green dots: F3 Orange dots: Fs Red dots: F2 (In example graph, F2 tests were not run.)
75	Power of MAAONVA Limited use of Preprocessing techniques Per gene estimation of factors contributing to variance permits random effects Systematic detection of treatment effects But: parametric model
76	LIMMA: Linear Models For Microarray Data Similar to MAANOVA Linear model of factors (treatment, dye, etc.) Factors are independent (no interaction) Requires a Contrast Matrix Uses either: log-odds (B) – Statistics (large B à DE gene) Bg=log (odds ratio) =l og (odds of gene g to be DE versus EE) moderated t-statistics (large T à DE gene)
77	How to define the Rejection Region: the multiple Hypothesis Problem
78	Test Of A Single Hypothesis gene g is equivalently expressed (EE) gene g is differentially expressed (DE)
79
80
81
82
83
84
85
86	Comparison of The Four Methods
87
88