Logo FlyChip
FlyChip
Functional Genomics for Drosophila
Cambridge Systems Biology Centre, Tennis Court Road, Cambridge, CB2 1QR, UK  [map]
Tel: +44 (0)1223-760280.   Fax: +44 (0)1223-760241.

Overview

  1. General
  2. Files overview
  3. Project info file: Pnnnnn_info.txt
  4. What images have you sent to me?
  5. Raw data
  6. Normalisation
  7. Result files
  8. Deciding which genes are differentially expressed
  9. Quality control & Diagnostic plots
  10. Useful analysis and visualisation tools
  11. Glossary of terms

1. General

Each project is split into replica groups that relate to pairs of samples. Each of these replica groups is hybridised to four microarrays (depending on how many arrays were initially requested). Two of these hybridisations are so-called dye swaps where a sample originally labelled with Cy3 is now labelled with Cy5 and vice versa. Dye swapping compensates for the differential labelling efficiencies of each dye.

Project info files

Describe how the replicate hybridisations are grouped together and how each sampled was labelled: which dye each sample was labelled with, which sample it was paired with and which slide it was hybridised to.

Projects codes

Use the format Pnnnnn, where nnnnn is the assigned project number.

Replica group codes

Use the format Rnnnnn. Some file names are based on which of the microarray slides they represent.

Microarray slide codes

Use the format Snnnnnn. Microarrays are numbered in the order they were printed..

Each of the terms and codes defined above will be used throughout the following sections. Most of the files we have sent to you will have names derived from these codes. We will now explain what each file type contains. This might be easier to understand if you simultaneously open the relevant file. To help with interpretation we include an example project, providing additional information for each section.


2. Files overview

This is a list of the files we have sent to you.


3. Project info file: Pnnnnn_info.txt

Pnnnnn_info.txt files describe how we processed your project and records which dye (or “channel”) was used to label each of your samples within each replica group. You may want to refer to this file when analysing your results.

N.B. Pnnnnn_info files are tab-delimited. They are best viewed using the freely available spreadsheets from OpenOffice or Microsoft Excel. Some text editors will have problems with the variable column spacing.

Caution - when using Excel be sure to use the open file command and indicate that columns such as gene names and unique IDs are labelled as text. If you do not do this, Excel may mangle some of these identifiers (PubMed).

Important note; dye swaps and gene expression ratios:

Project info file example (male vs female):

The following table describes project P99934, where we compared the gene expression patterns of adult male and female flies. This project comprises one replicate group (R10342) with four slides, two of which were dye swapped (S106438 and S106430). For this example, the ratio is male over female with the swap status 0 representing the male sample in the Cy5 channel and the female sample in Cy3 channel.

Project_Number Replicate_Group Slide_Number Hyb_Number Cy3_Image Cy5_Image Cy3_Sample_Name Cy5_Sample_Name Swap_status Comments
P99934 R10342 S106438 H10001 S106438_532 S106438_635 male_1 female_1 1 -
P99934 R10342 S106404 H10001 S106404_532 S106404_635 female_2 male_2 0 local background on bottom of slide
P99934 R10342 S106472 H10001 S106472_532 S106472_635 female_3 male_3 0 -
P99934 R10342 S106430 H10001 S106430_532 S106430_635 male_4 female_4 1 -

4. What images have you sent to me?

We provide you with two different types of images:

(1) Raw grey-scale 16-bit TIF images used to quantify how much labelled sample has bound to each spot for each channel. These files can be found in the images_bw folder. These are the primary raw data for your microarray experiment!

(2) False colour PNG images (color_images folder). These provide you with a visual and non-normalised representation of your results. These images can be used to check for slide-specific problems as well as for presentations.


5. Raw data

For each slide, spot-finding and quantitation is presently performed by dapple. The file name is based on which microarray slide it represents. This file contains the raw unprocessed data without taking any dye swap into account. To normalise these data yourself, you will need to consider the dye swap status. In this context, we recommend using the single channel data. For technical information on spot finding and quantitation tool, see Buhler J. et al (2000). Dapple: improved techniques for finding spots on DNA microarrays. University of Washington technical report, UWTR 2000-08-05 (report).

Header information is always denoted by a hash (#) at the beginning of the line. All other columns are defined below. However, please note that the first number within the grid_x column is the total spot number (e.g., 18240 for FL002), and the first number within the grid_y column is the total number of channels (e.g., 2).

Column definitions for Snnnnnn.state.dat

These first few columns denote the spot location in the microarray. Locations are provided using a system of Cartesian co-ordinates. The x-axis corresponds to the width of the image (the shortest side) and the y-axis corresponds to the length of the image (the longest side). The reference point for these co-ordinates (0,0) is the top left spot in each image.

The following columns describe the nature of each spot. The description includes the Drosophila transcript and the predicted gene for each spot. The last column defines whether the spot should be included in any normalisation if you should choose to do this yourself.

Note that genome annotations change and for the most up to date view of the gene models and transcripts interrogated by the INDAC oligonucleotide probes you can consult FlyMine at www.flymine.org

Further columns provide details about the spot status, signal and a pixel count for the foreground (i.e. the spot) and background (i.e. the area surrounding the spot). Spots with very few pixels in the foreground are probably unreliable because they contain too few pixels for any reliable spot signal estimate. Column headers N (or n) indicates the channel, with N=1 representing the Cy3 and N=2 the Cy5 channel.

What does this mean for the example of the male vs female comparison?

The raw data file of slide S106404 (S106404.state.dat) contains the foreground median intensity of the female sample in the fgMedian1 column (Cy3) and the male sample in the fgMedian2 column (Cy5). However, for the dye swap slide S106438, the male sample is given in the fgMedian1 column and the female sample in the fgMedian2 column. Therefore always check the swap status before analysing the raw data.


6. Normalisation

Measured fluorescent spot signals will differ systematically between different microarray hybridisations and dyes, including differences in background fluorescence, and in overall brightness with, e.g., one dye being twice as bright as another one. The process of correcting for such systematic differences is called normalisation. We use no background correction method, since this may significantly increase the overall variance of the data.

Differential expression is presented as a ratio of Cy5 over Cy3. A more symmetric (i.e., a “Gaussian” or “normal-like”) distribution is achieved using log-transformation with a log[2] scale.

The log differential expression ratio for each spot or 'M' is calculated as follows:

M = log2 (Cy5 / Cy3) or M = log Cy5 - log Cy3

The log intensity of the spot or 'A' (a measure of the overall brightness of the spot) is:

A = (log2 (Cy5 * Cy3) /2) or A = ( log Cy5 + log Cy3) / 2

In our files, M-values represent:

Depending on the quality of the data, three different methods of data normalisation are available. Our prefered normalisation method uses the vsn software and is performed using tools from Bioconductor.

Variance Stabalisation Normalisation (vsn) - our default normalisation method

This method is based on the work published by Huber et al. (2002). Bioinformatics 18(1), S96-104 and the corresponding Bioconductor package is called vsn.

For each dye and microarray, the background fluorescence and a factor reflecting overall brightness are inferred to make the signals identical for this subset of non-differentially expressed genes. A necessary assumption is that more than half the genes are NOT differentially expressed. For further (technical) information, see Huber et al. (2002) Bioinformatics 18(1), S96-104 (abstract).

Loess & Quantile normalisation (lquant)

Performs a within-array global loess normalisation of the M-values, followed by a between-array quantile normalisation. These normalisation methods are implemented in the Bioconductor limma software package.

Loess normalisation assumes that the bulk of probes on the array are NOT differentially expressed. It does not assume that there are equal numbers of up and down regulated genes or that differential expression is symmetric about zero. For further information, see Yang, Dudoit et al. (2002) Nucleic Acids Res. 30(4), e15 (abstract).

The aim of quantile normalisation is to ensure that all the intensity distributions on each array are identical. It involves an initial array-specific centering of the data, with the centred data being subsequently ordered from lowest to highest. Afterwards a distribution is calculated whereby the lowest value is the average of the lowest expressed gene on each of the arrays. This calculation is repeated for each subsequent order of intensity values up to the average value of the highs from each of the arrays. Each measurement on each array is then replaced with the corresponding average value in the distribution. For further (technical) information, see Bolstand et al. (2003) Bioinformatics 19(2), S185-93 (abstract).

Quantile normalisation (quant)

Performs a between-array quantile normalisation that does not correct for within slide bias. This normalisation method is implemented in the Bioconductor limma software package.


7. Result files

We perform a very basic statistical analysis of your data using different analysis tools. This includes a p-value estimation of your data for direct comparison, e.g., sample versus control. For other experimental designs, e.g., reference designs, or time-course analysis, we do not routinely perform a statistical analysis. However, we offer a cost-recovery based analysis service. We have listed some analysis tools in section 10 that can be used for these designs.

The three statistical analysis tools we use differ in their stringency of assigning significance to the data, with CyberT being the least stringent (gives the most genes with p-value < 0.05). Limma and siggnes are the most stringent (fewest genes with a p-value < 0.05).

Result file names consist of the replicate group number, normalisation method identifier and the statistical software used for the analysis. For each replicate group of your project we produce summary files which describe the nature of each spot, the transformed normalised intensity differences between the Cy5 and Cy3 channels for the replicate slides (M-values), and the transformed normalised average intensities of the replicate slides (A-values). Normalised data have the dye-swap taken into account; therefore, all M-values should be treated as if there had been no dye swap.

Tip: Excel will change cell contents, e.g., some of the genes names! To avoid this, always open excel first, and then change text columns to “text format” while importing the data (PubMed).

All result files have the following columns in common. The first few columns contain the gene information.

The following columns contain the M and A-values for each slide within this replicate group. Positive M-values indicate an increase in relative intensity (Cy5 greater than Cy3), negative values indicate a decrease in relative intensity (Cy5 less than Cy3). Remember that both M and A-values are log2 transformed. Numbers of equal value but opposite sign indicate equivalent fold changes up and down respectively. N.B. Since dye-swaps have been taken into account, the numbers across all replicate slides are comparable and should ideally change in the same direction.

The next column indicates whether a spot was saturated during the scanning process. The scanner has a pixel intensity range from 1 (20) to 65535 (216 -1). Image brightness can be changed by choosing a different PMT gain. We try to reduce the amount of saturated spots, however, there is a trade-off between having few saturated spots and losing low intensity spots, when scanning at lower PMT gains. Spots with raw intensities >65000 in either one or both channels are flagged with “1” in the SaturationFlag column.

The next column provides an indication of the spot quality. This can be used to determine whether the reported expression changes are reliable or subject to error due to problems with the either the printing, hybridisation, or spot-finding.

The following columns are specific for each analysis tool.

Rnnnnn.vsn.CyberT.tab

CyberT is a statistics program with a web interface that can be conveniently used on high-dimensional array data for the identification of statistically significant differentially expressed genes. It employs statistical analyses based on simple t-tests that use the observed variance of replicate gene measurements across replicate experiments. We have implemented the R code hdarray in our analysis pipeline. For further technical information, go to the CyberT web page (http://cybert.microarray.ics.uci.edu/). The first few columns are as described above, followed by the CyberT analysis specific columns.

Rnnnnn.vsn.Limma.fdr.tab

Limma stands for Linear Models for Microarray Data. For each gene, it fits a linear model to the expression data and employs an empirical Bayes method to stabilise the analysis.

A design matrix and a contrast matrix have to be established for the data of a given replicate group. In a paired-data design the number of coefficients is one fewer than RNA sources (e.g. wildtype vs mutant, the number of coefficients equals 1). The first step is to fit a linear model that describes the systematic part of the data, followed by moderated t-statistics calculated for each probe and each contrast, respectively. Moderated t-statistics allow the same interpretation as an ordinary t-statistic although the standard errors have been moderated across genes, i.e. lowered towards a common value by using a simple Bayesian model. P-values are then adjusted using Benjamini and Hochberg's step-up method for controlling the false discovery rate (FDR). For more information, see Smyth (2004) Statistical Applications in Genetics and Molecular Biology, 3:Article 3 (abstract).

Again the first few columns are as described above followed by the limma analysis specific columns.

Rnnnnn.vsn.siggenes.delta.nn.tab

We can perform the popular Significant Analysis of Microarrays (SAM) using the siggenes software package in Bioconductor. SAM identifies genes with statistically significant changes in expression by assimilating a set of gene-specific t-tests. Each gene is assigned a score on the basis of its change in gene expression relative to the standard deviation of repeated measurements for that gene. Genes with scores beyond a certain threshold are deemed potentially significant. The percentage of such genes identified by chance is the false discovery rate (FDR). To estimate the FDR, nonsense genes are identified by analyzing permutations of the measurements. The threshold can be adjusted to identify smaller or larger sets of genes, and FDRs are calculated for each set. For further information, see Tusher et al. (2001) PNAS 98(9), S5116-21 (abstract).

The analysis for a replicate group creates a table (Rnnnnn.vsn.Siggenes.mat.fdr.table.txt), stating the number of called genes, false genes and the FDR for the delta values. Ideally, we choose the delta value that has a FDR < 1 and a number of false genes < 1. However, in some cases it might be appropriate to include a larger number of genes, particularly, if there are only very few genes in the list. Please refer to this table for the FDR cutoff and number of predicted false genes for your results.

Again the first few columns are as described above followed by the limma analysis specific columns.

Pnnnnn.vsn.matrix.txt

For your convenience we have combined the data of all replicate groups of your project including the (adjusted) p-values of all three analysis tools within one file. This file facilitates a comparison between the different analysis tools and replicate groups. The replicate group number has been added to the column header to clarify what replicate group M-, A- values or p-values belong to. The averageM column has been omitted as it makes no sense to average over several replicate groups.

Pnnnnn.saturated.matrix.txt

Saturated spots (raw intensity value > 65000) are flagged in the result files. If a gene is saturated in one channel but not in another the ratio is compressed. To facilitate reviewing which genes were saturated in which channel and on which slide, we have created the Pnnnnn.saturated.matrix.txt which lists all spots which have been flagged.

The first few columns contain again the gene information, then the raw intensity values for each slide are listed.

Please note that the dye-swap has not been taken in consideration.

Male vs female example files

Below are extracts of result files of the male vs female example. The ratio defined in the Pnnnnn_info.txt is male/female. After sorting the files by the averageM column, the female specific genes are displayed at the top of the list. Yp2 (Yolk protein 2) has an averageM value of -8, this means that this gene is 256-fold higher expressed in females than males in this particular experiment. For Yp3 you can see that at least one of the channels was saturated on all 4 slides.

Siggenes analysis creates two files:

Summary file of the project:

There were two saturated spots in this project:


8. Deciding which genes are differentially expressed

Since we do not rank the data, there are several points to consider before deciding whether a gene in your experiment is differentially expressed.


9. Quality control & Diagnostic plots

All RNA samples of your project have to pass a RNA quality control test. Subsequently, the good quality RNA samples are mixed with spike control RNA in a reverse transcription or amplification reaction and labelled. After the hybridisation the signals of these spike control spots are checked to verify a successful labelling and hybridisation. The quality of our printed arrays within each print batch is ascertained by staining a random sample of slides (protocol). We perform quality control of the data before and after the normalisation and are able to identify problem slides. Any problem we identify will be reported in the corresponding comment column in the project index file. Depending on the severity of the problem we will occasionally exclude a slide from the analysis (this will be stated in the project index file). Minor problems (e.g. high local background level on a slide) can be solved by rejecting spots in the affected array area and will result in a RR-flag for these genes in your file.

The following examples were primarily taken from the aforementioned male versus female example. The plots were created using the following packages in Bioconductor: limma, vsn, marray and siggenes. For reasons of brevity, not all images for all slides have been displayed.

Background and Foreground images

FL002 arrays were printed using a print head with 4x12 arrangement of print-tips and the microarrays are partitioned into a 4x12 grid of tip groups (FL003 design). Each grid was printed with a single print-tip. It is interesting to look at the variation of background and foreground intensity values across the array. For each slide we create three images displaying, from left to right, the log2 transformed raw intensity values of the green (Cy3), red minus green and red (Cy5) channel. The darker the colour, the higher the signal intensity. In the background images slide S108122 shows that there is high background mostly on the left side of the array. The foreground plots indicate that this background will cause problems when looking at the ratio (red minus green). We reject spots during the spot finding in areas with high background, theses spots will have RR-flags in the result files.

Rnnnnn.raw.Backgound.png
Rnnnnn.raw.Foregound.png

Scatter plots with correlation coefficients

The following plots are created for the raw data as well as the normalised data. Only spots with good quality flags on all slides are included, spots that contain RR-flags in any one slide are removed. In the upper right panels, pair-wise scatter plots of the raw intensity values of each channel within a replicate group are shown. The lower panels show the correlation coefficients. Channels which contain the same sample type should be close to the red line and have a correlation coefficient close to 1. Below is the raw scatter plot for the male vs female example, showing that there are large differences between the samples. Channels S106438.Cy3 and S106472.Cy5 both contain male samples.

R10342.raw.scatterplots.png

MA plots

MA-plots, where M is the log differential expression ratio (i.e., the expression ratio) and A is the mean log intensity between the two channels of the slide (i.e., the spot signal). In the raw plots control spots are colour coded (see table below), bad quality spots [RR or RS-flagged] are not plotted and low intensity spots show larger variance (left side of the plot). The data appears symmetrical around the horizontal line if the two channels behave similarly. After normalisation only gene spots are plotted (control spots are omitted). The example below shows two dye-swap slides, therefore the M-values (ratios) of the second slide are a mirror image of the first slide. The MA-plot of slide S106404 after normalisation could indicate that there is a problem with this slide, however if one ignores the low intensity spots (A-values below 5) the plot looks similar to S106438.

Spot Color Spot Type
Black Gene spots
Green Controls (e.g. degradation probes, FLPase, LacZ, modified_GFP, Gal4_cds)
Orange Spikes
Yellow Empty
Red Spotting Buffer

R10342.raw.MAplots.png

After normalisation only gene spots are plotted (control and bad quality spots are omitted). The red dotted line shows the loess fit. The MA-plot of slide S106404 after normalisation could indicate that there is a problem with this slide, however if one ignores the low intensity spots (A-values below 5) the plot looks similar to the first slide.

R10342.vsn.MAplots.png

MAprinttip plots

Raw data MA-plots, with colour-coded lowess fits for each print-tip-groups. These plots can highlight printing or hybridisation artefacts of the array. Ideally all lines should be close to the zero line.

R10342.raw.MAprintipplots.png

Boxplot plots

Boxplots of the M-distribution per print-tip-group can be useful to identify spot or hybridisation artefacts. The central box in the plot represents the inter-quartile range (IQR), which is defined as the difference between the 75th percentile and 25th percentile, i.e., the upper and lower quartiles. The line in the middle of the box represents the 50th percentile, i.e., the median. Extreme values, greater then 1.5 IQR above the 75th percentile and less than 1.5 IQR below the 25th percentile, are typically plotted as individual data points.

R10342.raw.Boxplots.png

Additionally, we create box plots of the overall M-distribution for each slide of the raw and normalised data. After normalisation the median should be centred around zero and the spread should be similar between the slides.

R10342.vsn.Boxplot.png

Density plots

Density plots display smoothed empirical densities for the individual green and red channels. Without any normalisation there is considerable variation between both channels and between arrays.

R10342.raw.Densityplots.png

Summary plots of the densities of all slides within this replicate group before and after normalisation.

R10342.vsn.Densityplots.png

Mhist and Ahist plots

The histograms for the M- and A-values should be similar for slides within a replicate group. In the Ahist-plot of slide S106404, as in the MA-plot of this slide, some of the low intensity spots are causing a distortion visible as a peak to the left of the graph.

R10342.vsn.Mhist.png

R10342.vsn.Ahist.png

Sample Cluster Dendogram

All samples are normalised together and cluster dendograms are calculated using the euclidian distance method. Identical sample pairs (biological repliate samples) should cluster together, allowing outliers to be detected. However, if the differences between the different sample types within the project are very small the Dye effect may be stronger than the sample difference, so that the samples cluster by Dye.

P99934.sample.cluster.png

Standard deviation versus rank of the mean and the mean plot (only for vsn normalised data)

The aim of these plots is to determine whether there is a systematic trend in the standard deviation of the data as a function of overall expression. The assumption that underlies the usefulness of these plots is that most genes are not differentially expressed, consequently the running median should be a reasonable estimator of the standard deviation of feature level data conditional on the mean. The red dots show the running median of the standard deviation. The curve given by the red line is an estimate of the systematic dependence of the standard deviation on the mean. After Variance Stabilization Normalization, this should be a horizontal line. It may have some random fluctuations, but should not show an overall trend. If this is not the case, that usually indicates a data quality problem, or is a consequence of inadequate prior data pre-processing. The rank ordering distributes the data evenly along the x-axis.

R10342.vsn.stdev_vs_mean.png

Limma Volcano plot

A volcano-plot of log-fold-change versus a non-negative statistic that tests the significance of the fold change. Since the size of the statistic tends to increase with absolute log-fold change, such a plot has the characteristic shape of the open crater of a volcano. Volcano plots are sometimes used to emphasise that variability plays a role in significance as well as fold-change. Coloured dots represent probes that show large and very significant fold changes (adjusted p-value < 0.05).

R10342.vsn.Limma.fdr.png

Siggenes plot

Visualisation of SAM plot using the in the analysis specified delta value specified in the analysis. The green dots in the distribution represent the up (top) and down-regulated (bottom) genes.

R10342.vsn.Siggenes.delta1.5.png

10. Useful analysis and visualisation tools

There are many tools available for normalisation and statistical analysis of microarray data. See data analysis for an introduction to data analysis and links to useful tools.

Data analysis depends on the experimental design and will be different for each project. However, to get you started we have listed below links that might be useful. Please refer to the tool specific software documentation for details.

p-value calculation

Identify differentially expressed genes and estimate False Discovery Rates.

Tool Type Direct design Time-course design Reference design
CyberT online tool yes - yes
GEPAS online tool yes yes yes
limma bioconductor tool yes yes yes
siggenes bioconductor tool yes yes yes
maSigPro bioconductor tool - yes -

Clustering

Identify genes with similar expression profiles using clustering.

Tool Type
Cluster 3.0 & TreeView free download tools
maSigPro bioconductor tool
GEPAS online tool

11. Glossary of terms

Version 2.0. B.Fischer (12-03-2008)