Low-level Analysis. cdna Microarrays. Lecture 2 Low Level Gene Expression Data Analysis. Stat 697K, CS 691K, Microbio 690K

Transcription

1 Lecture 2 Low Level Gene Expression Data nalysis Stat 697K, CS 691K, icrobio 690K Statistical Challenges odel variation of data not related to gene expression Compare expression for the same gene across different experiments Identify genes that show different expression patterns in case and control Cluster genes that show similar expression Classify genes into different cancer tumor types: Invasive Non-invasive 2 Low-level nalysis cdn icroarrays Pre-processing steps Image analysis Normalization Remove systematic biases due to experimental artifacts, not related to true gene expression 3 4 1

2 cdn microarray slides Spotting area = 72 mm x 22 mm Slide dimension = 76 mm x 25 mm x 1 mm Scanning icroarray images are produced by a laser scanner The scanner performs an area scan of a slide Produces a digital map, or image, of the fluorescent intensities for each pixel The scanner produces two files, one for each fluorescent dye (each file is a 16-bit TIFF) (3 x 1 x.04 inches) (standard glass slide) 5 Yang et al 2002 JCGS 6 Scanning Different dyes absorb and emit light at different ranges of wavelengths To measure abundance of the 2 fluorescent dyes for each spot The scanners generate excitation light at different wavelengths and detect different emission wavelengths The commonly used cyanine dyes Cy3 (green) and Cy5 (red) have emission in the nm (nanometer) and nm ranges. The 2 images are overlayed for visualization purposes: Called RGB overlay image -- Yellow = nondifferentially expressed -- Red = Cy5 -- Green = Cy3 Yang et al 2002 JCGS 7 8 2

3 Image nalysis The goal of image analysis is to extract measures of the red and green fluorescence intensities (= transcript abundance) for each spot on the array. Intensities for each pixel range from 0 to (2 16-1) (=65,535)) Have a potentially large impact on subsequent analysis: determining differentially expressed genes clustering genes Software for Image nalysis Dapple (Buhler, 2002) GenePix (Eisen, 1999) ImaGene (Eisen, 1999) Scanlyze (Eisen, 1999) Spot (Yang and Buckley, 2001, built on R) Quantrray (Packard Bioscience [co]) Derray (Scanlytics [co]) 9 10 Image nalysis 3 steps Gridding: estimate location of spot centers Segmentation: classify pixels as foreground (signal) or background Information extraction: for each spot and each channel, calculate Signal intensities Background intensities Quality measures Gridding Example array: 4 rows, 4 columns of grids within each grid there are 19 rows, 21 columns of spots Gridding: assigns coordinates to each spot separates rows and columns of grids 11 ost software systems allow for both manual and automatic gridding procedures 12 3

4 Segmentation Goal of segmentation: classifying pixels of a spot as either foreground (signal) or background foreground background 4 ethods of Segmentation 1) Fixed circle segmentation 2) daptive circle segmentation 3) daptive shape segmentation 4) Histogram segmentation Fluorescent intensities can then be calculated for each spot (take mean, median of pixels) as measures of transcript abundance ) Fixed Circle 2) daptive Circle Fits a circle with constant diameter to all spots in the image Easy to implement Not good if shapes and sizes of spots vary First seen in Scanlyze, also in GenePix and Quantrray, and is an option in most packages Yang et al., The circle s diameter is estimated separately for each spot. Not good for non-commercial arrayers due to less than perfect circular shape of spots due to: Print tip problems Uneven solute deposit Insufficient time for rehydration after slide printing Implemented automatically in Dapple User option in GenePix to adjust each spot: time consuming Yang et al., 2001 daptive circle estimates each diameter separately 16 4

5 3) daptive Shape Segmentation ethod: Seeded Region Growth (SRG) Specify starting pixels, or seeds, for both foreground and background for each spot. The algorithm assigns neighboring pixels to either foreground or background based on intensities. dvantage for microarrays: -- know exact number of spots -- know approximate placement of spots SRG implemented in software Spot Not in widely used software packages foreground (all else is background) Results from SRG 4) Histogram Segmentation Uses a target mask: area larger than all spots (target site=cdn region target patch=square area surrounding each spot) Defines foreground and background as the mean intensities between some predefined percentile values. e.g. background=mean of 5%-20% foreground=mean of 80%-95% Implemented in Quantrray (variations in Derray) target site Chen et al., 1997 Yang et al., Summary of Segmentation ethods Fixed Circle ethod daptive Circle daptive Shape Histogram Segmentation Software or lgorithm Scanlyze, GenePix, Quantrray GenePix, Dapple Software: Spot lgorithm: region growing, watershed Software: ImaGene, Quantrray, Derray lgorithm: adaptive thresholding 19 Information Extraction Goal of information extraction: to calculate one value for foreground and background intensity, and a quality measure Foreground: Histogram segmentation method automatically calculates foreground (and background) intensity values Other methods: Foreground intensities: - ean of spot pixel intensities, or - edian of spot pixel intensities (some software reports both) Foreground: Take the average or median of pixel intensities Yang et al.,

6 Background Intensity Why do we measure background? - Some intensity measured on slide is not related to the sample: chemicals on glass, non-specific hybridization - Even a sample of water on the slide will register some intensity - The background value is a measure of intensity not related to the sample, and is subtracted from foreground to get a measure of intensity for the spot ** Estimation of background has a large impact on results Background Intensity ethods for calculating background intensities: ean of pixels surrounding the spot edian of pixels surrounding the spot Constant (global): common background for all spots None Yang et al., Yang et al., Defining Background Pixels Foreground Quantrray: background is everything between 2 green circles Scanlyze: background is everything inside blue box, outside foreground Comparison of Background Correction ethods See Yang et al JCGS No conclusive results Differs for each data set Paper offers some suggestions for specific types of images Spot : background is everything within 4 pink squares Yang et al.,

7 Calculating Expression For each spot on the slide we calculate: Red intensity = Red foreground Red background Green intensity = Green foreground Green background and combine them in the log (base 2) ratio Log 2 ( Red intensity / Green intensity) Calculating Expression Ways to calculate foreground background: 1) mean(foreground) mean(background) 2) median(foreground) median(background) If these values are less than zero for one channel, can assign a minimum positive intensity value (i.e. 10). If both red and green are less than zero, remove spot from analysis Why Log-Transform Data Log-transformation of Data Evens out highly skewed distributions Gives a more realistic sense of variation Logs base 2 are the easiest for work with (intensities are typically a number between 0 to 65535=2 16-1) interpretation is similar: positive and negative values (source: Terry Speed s group, UC Berkeley)

8 Histograms This is the most common graphical display of quantitative data. Histograms give information about the distribution of values for a numerical variable Range of values Frequency of values observed in a class, by grouping values into classes. Example: Distribution of log-ratios on a cdn microarray Source: Darlene Goldstein Some general histogram forms left-skewed right-skewed Quality Filtering symmetric Source: Darlene Goldstein 31 8

9 Comet Tails Likely caused by too slow immersion of the slides into a solution (succinic anhydride blocking solution) Practical Problems 1 Practical Problems 2 High Background One likely cause: Precipitation of the labeled probe. Results in weak signals Practical Problems 3 Practical Problems 4 Dust Spot overlap: Likely cause: too much rehydration during post - processing

10 Red/Green overlay images Overlaying red/green images offers a quick visualization of: - color balance - uniformity of hybridization - spot uniformity - background - artifacts such as dust or scratches Spatial plots: background from the two slides good Good: low background, many differentially expressed genes. Bad: high background, ghost spots, few differentially expressed genes 37 Low quality slide has non-uniform background distribution Plotted using Bioconductor packages for R bad 38 Quality easurements Produced by Image nalysis Software rray-wide quality Percentage of spots with no signals Range of intensities Distribution of spot areas (min, max, quartiles) Spot-specific quality Spot size (area in pixels) Identification of bad spots (spots with no signal, or flagged spots) Foreground / Background ratios don t want background > foreground Variation in pixel intensities within a spot Circularity measure (area/perimeter 2 as small as possible) 39 Normalization 10

11 Normalization Why? To correct for systematic differences between samples on the same slide, or between slides, which do not represent true biological variation between samples. Normalization: otivation rray Data biological variation experimental variation (interested) (not interested) How do we know it is necessary? By examining self-self hybridizations, where no true differential expression is occurring. -- self-self hybridizations label the same sample red and green and hybridize the mixed sample to the microarray -- result is not on a 45 degree line ** We find dye biases which vary with overall spot intensity, location on the array, plate origin, pins, Carefully control experimental steps (e.g. PCR amplification, slide printing, reverse transcription, sample labeling) Wish to minimize Normalization Normalization methods could have larger effect on analysis than the downstream steps such as group comparisons (Hoffmann et al Genome Biology) scanning parameters Normalization Identify and remove sources of systematic variation such as: Different labeling efficiencies of the dyes Different amounts of Cy3- and Cy5-labeled mrn Different scanning parameters Print-tip, spatial, or plate (96 well, 384 well, etc.) effects Where Normalization Fits In Sample Sample Preparation Preparation Hybridization rray rray Fabrication Fabrication Scanning + Image nalysis Normalization Data Data nalysis nalysis Spot location, assignment of intensities, background correction, etc. Normalization Subsequent analysis, e.g clustering, uncovering genetic networks

12 Types of Normalization Location Scale Within-slide Paired slides (dye swap) Self normalization cross multiple slides Within Slide Normalization Goal: to compare different experiments or replicate slides to each other There are different amounts of exposure/brightness for each slide Normalize within each slide to compare slides Within Slide Normalization Normalization balances red and green intensities Imbalances can be caused by Different incorporation of dyes Different scanning parameters In practice, we usually need to increase the red intensity to balance the green 47 0 Location Transformation log 2 (green/red) Large % differentially expressed, but it is a dye effect Small % differentially expressed ean & edian centering are examples of location transformations 48 12

13 Constant (Linear) Normalization Example: green dye has shown better labeling efficiency ssume red and green dyes are related by a constant factor, i.e., G=kR Normalize by dividing all ratios on slide by k. common choice for k is the median or mean of the log-intensity ratios on the entire slide Result: For k=median of all ratios, the resulting median of all ratios on the slide will be 1 (green=red) the resulting median of all log-ratios will be 0 (green=red) --Intuition: we expect most genes to be non-differentially expressed, so that green=red. --For example: approximately 5% are differentially expressed --fter normalization, can compare different slides to each other Example of Constant (Straight Line) Normalization (location transformation on a log scale) Gene 1 Gene 2 Green Red G/R 4 4 Log 2 (G/R) 2 2 Log 2 (G/R)- median (Log 2 (G/R) 0 0 eans vs. edians ean = sum(all values)/total number edian: sort values, take middle. edians are preferred since they are robust to outliers, i.e. are not influenced by unusually large or small values. Gene Gene edian

14 eans vs. edians 5 values, mean and median similar mean=21 median=20 5 values with an outlier mean=215 median=20 Non-constant normalization Previous method used a constant value to normalize But relationship between red and green is typically not constant (i.e. not linear) Here, mean is not a good representation of the data Data Visualization If perfect comparability of red, green: straight line through 0, slope 1 (45 degree line) Use - Plots Instead When we have two sets of numbers such as R and G varying over a large range: it is useful to compare log R with log G by plotting their difference: =log(r/g) vs. their average: =(1/2)log(R*G) Proposed by Dudoit et al Doing this we might see something unexpected By contrast, plotting log R against log G is typically much less revealing can give a quite unrealistic sense of concordance 55 Source: Terry Speed 56 14

15 lways Log, lways Rotate Log(Red) vs Log(Green) - Plot - Plots - plot is 45 rotation of standard scatter plot log R 45 log G = log (R/G) = ½[ log R + log G ] log 2 R vs log 2 G =log 2 R/G vs =(1/2)log 2 RG plots Idea: look at log-ratios of expression as a function of overall intensity Genes with very small intensity in one channel (red) and large intensity in another channel (green) should be examined carefully = log 2 (R/G) - Before Normalization Ratios depend on average intensity, so normalize using an average dependent normalization 59 = log 2 (R*G) / 2 - ost log-ratios above 0 - Should not use constant (straight-line) normalization 60 15

16 Lowess: intensity-dependent normalization Non-constant, non-linear variation Variability is non-linear due to: spatial heterogeneity problems correcting for background Experimental variation such as: separate reverse transcription of red/green samples separate labeling of red/green samples Use the scatter plot smoother Lowess : local average-intensity dependent normalization Lowess Smoothing LOcally WEighted Scatterplot Smoothing robust, local smoother of scatter plot data Uses robust locally linear fits Steps: window is placed about each x value; User can determine width of window Points inside window are weighted so that points nearer to x get most weight Lowess is built into R Reference For microarray data: Yang et al, Nucl. cid. Res. (2002) Yang et al Un-normalized - Plots Normalized Lowess Demo Normalized values are just heights between spots and the general trend (red line) The red line is found by Lowess

17 Lowess Demo 2 Lowess Demo 3 Choose x value, choose window, calculate local linear regression within window Repeat for each value of x Lowess Demo 4 Lowess Demo 5 Smaller windows yield a smoother curve Widths Too Narrow

18 Lowess Demo 6 Lowess Demo 7 Widths Too Wide Final lowess curve Lowess Demo 8 Lowess Demo Span f Subtract the lowess value from each value to get the normalized curve Span f = fraction of data used at each point 20% 40% (R option) Window width depends on f

19 Lowess procedures Before Lowess Once the Lowess curve is found, subtract the Lowess curve from every value, for every x. fter Lowess 73 The distribution centers around 0 after Lowess 74 Normalization Result The distribution centers around 0 after normalization Outliers are differentially expressed (either upregulated or down-regulated genes) Print-tip Group Normalization

20 Print-tip Group Normalization Each block on the array was created using a separate print-tip (i.e. pin) Total of 16 print-tip groups Normalize data from each print-tip group separately using Lowess Print-tip Group Normalization Dudoit et al Yang et al Lowess curves, one for each print-tip group Dudoit et al Yang et al Print-Tip Group Normalization Print-Tip-Group Normalization Before fter Dudoit et al 2002 Yang et al Before fter Yang et al

21 Boxplots boxplot provides a visual summary of the distribution of values for a numerical variable ore detailed than a histogram Shows median of data and outliers Quantiles The p th quantile is the number that has the proportion p of the data values smaller than it 30% Source: Darlene Goldstein 81 Source: Darlene Goldstein 5.53 = 30 th percentile 82 easures of Spread Inter-Quartile Range: IQR The 25 th (Q 1 ), 50 th (median), and 75 th (Q 3 ) percentiles divide the data into 4 equal parts; these special percentiles are called quartiles Theinterquartile range (IQR) of a variable is the distance between Q 1 and Q 3 : Five-number summary and boxplot n overall summary of the distribution of a variable is given by five values: in, Q 1, edian(q 2 ), Q 3, and ax boxplot provides a visual summary of this five- number summary Whiskers = 1.5 x IQR (1.5 x IQR from the box, or to max and min, if IQR = Q 3 Q 1 less) **50% of the data falls in the interquartile range Source: Darlene Goldstein 83 Points outside whiskers are outliers Source: Darlene Goldstein 84 21

22 Example suspected outliers Print-Tip-Group Normalization Q 3 median `whiskers Q 1 Source: Darlene Goldstein 85 Before fter Yang et al ultiple Slides Normalization ethod Goal: to combine information across multiple (or replicate) slides Need to normalize across slides Extension of within-slide normalization Scale normalization step maybe skipped if differences between slides are small Trade-off between the gains achieved by scale normalization and the possible increase in variability introduced (we will go into detail in ffymetrix data lectures) 87 Within-slide Normalization Results for 3 Replicate Self-Self Hybridizations Before Normalization fter Within Print-tip Group Normalization Can combine data across 3 slides after within-slide normalization djusting for scale may introduce more noise 88 22

23 Normalization by Experimental Setup 1) Dye swaps: reverse green and red dyes between samples 2) Spot specific genes multiple times on array; controls for spatial effects Normalize with House-Keeping Genes Genes that perform routine cell functions Constantly expressed across samples and experiments Draw Lowess curve through house-keeping genes Potential problems The number of predetermined available house-keeping genes may be small Their intensities do not cover the entire range of intensity values (most of them in the high end). Not representative of all genes The expression levels of house-keeping genes can exhibit natural variability (Novak et al., Genome Biology 2002) Normalize using Rank-Invariant Set of Genes From Wong group, see Schadt et al. 2001, Tseng et al. 2001, Li and Wong 2001b lso Stuart et al (we will go into detail in ffymetrix data lectures) Credit These slides are based in large part on lectures by Steve Qin, University of ichigan, with generous permission. Steve Qin Cheng Li Jun Liu Wing Wong Yee Hwa Yang Sandrine Dudoit Percy Luu Terry Speed Debashis Ghosh Jeff Townsend Ryan Baugh Rafael Irizarry David Hoyle, University of anchester oarray Darlene Goldstein