Micro array chip contain multiple probes for each gene. After completion of the experiment a .dat is generated which is a image showing the intensity of each probe in a digital format. Now it is converted into .cell file which contains the real probe intensity numerical values.The higher the probe intensity higher is the expression of the gene. But its meaningless until you compare it with the intensity of same probe in control sample. These probe intensity values are preprocessed (Background correction, normalization, Perfect match correction and Summarization) using diff algorithms (RMA, Mas5). The normalized probe intensity values are further used for calculation of relative gene expression i.e treatment vs control to get the fold change. The positive and negative values indicates up-regulated and down-regulated genes in treatment group when compared to control.

The raw data is effectively unitless as it is simply relative signal intensity. Flourescence is measured by a photo multiplier tube or charge-coupled device and signal scaled across the range of detection for the platform. The starting data is thus simple relative signal values without units, based effectively on simple counted photons by the PMT or CCD.

Thanks for a rapid response! This means that when one sample has 5 times more expression than the baseline, and another sample is twice as low as the baseline, I cannot infer any true amount of expression between the two? I think his main question is how much is 1.0 in the scale of microarray detector in terms of relative expression...

Correct, a microarray experiment does not and cannot measure absolute expression, only relative expression. So any single measure is meaningless - it can only be interpreted relative to another measurement. The scale only applies to that specific array and it's upper and lower bounds. That is why to compare multiple arrays one must normalize them to compare the normalized expression values.

Micro array chip contain multiple probes for each gene. After completion of the experiment a .dat is generated which is a image showing the intensity of each probe in a digital format. Now it is converted into .cell file which contains the real probe intensity numerical values.The higher the probe intensity higher is the expression of the gene. But its meaningless until you compare it with the intensity of same probe in control sample. These probe intensity values are preprocessed (Background correction, normalization, Perfect match correction and Summarization) using diff algorithms (RMA, Mas5). The normalized probe intensity values are further used for calculation of relative gene expression i.e treatment vs control to get the fold change. The positive and negative values indicates up-regulated and down-regulated genes in treatment group when compared to control.

This is a huge URL, but the PDF is actually pretty informative, albeit a bit old. It covers the actual physical detection of signal in array experiments

http://www.moleculardevices.com/Documents/general-documents/mkt-appnotes/microarray-appnotes/Understanding_Fluorescent_Microarray_Imaging_Instruments.pdf

Keep in mind that microarray technology covers a lot more than merely Affymetrix technology and the GEO database contains data from a range of microarray technologies - Biochips, Agilent arrays, Affy arrays, spotted arrays, and so on. They all just rely at some point on the counting of photons of emitted fluorescence.

Often (usually hopefully) the data in GEO includes two distinct entries per experiment - the raw signal intensity data, and the data as normalized for comparative analysis by the submitter. The raw data for an Affymetrix experiment would be a *.CEL file, and is provided by submitters so you can reanalyze the data as necessary for your own specific comparative analyses. Ordinarily, the first step of any microarray analysis is to normalize all the raw data to be included for comparison.

I can do some nitpicking on Michaels comment: the measured values are light (fluorescence) intensities. They do have a unit, typically V [Volts] because it is measured as a voltage. This unit is rarely interesting, and it is amplified and digitized in integer numbers of a particular size (mostly 16bit). Since this is a simple monotone transformation (with some rounding), these numbers actually still have a unit. It thus would be most correct to talk about "arbitrary units" [AU] of fluorescence intensity.

Typically the measurements are presented as logarithms. Logarithms can only be taken from unitless numbers, so actually the logarithms are the logs of the ratio of the fluorescence intensity [AU] and the unit fluorescence intensity [AU]. So when the measured signal is 1024 AU, the log2 is actually calculated as log2(1024 AU / 1 AU) = log2(1024) = 10, what is a unitless/dimensionless number (the unit cancelled out).

Differences between such log-Intensities represent log-ratios of intensities. When for instance the log2-Intensity under control conditions is 10 and under treatment conditions is 7, the difference is 7-10 = -3, indicating that the measured intensity under treatment conditions is 2^(-3)-fold of the intensity of the controls (= one eigth), or the other way around: the intensity in the controls is 2^(+3) = 8 times as high as under treatment conditions.

Note that the logs nicely represent the symmetry of fold-changes: an x-fold induction and an x-fold repression would numerically be the same, just with opposite sign. The logs have some further nice properties that make them attractive for a statistical analysis.

I suppose it may also be worthwhile to mention that the "raw" end-user data is not the actual raw scan data either. On, for example, an Affymetrix scanner, the instrument software (Command Console) handles the actual image processing to generate the CEL file written out for analysis. A CEL file represents already summarized image data - it includes the summary intensity value for each feature on the array (as well as the standard deviation of that intensity and the number of pixels interrogated to derive that feature summary value), but the CEL files do not contain the actually scan data for each pixel interrogated. The pixel intensities would be in the *.DAT file, which is a binary data file.

At least for Affy arrays, if you want to see the details of what is in a CEL file, Affy makes all that information readily available - http://media.affymetrix.com/support/developer/powertools/changelog/gcos-agcc/cel.html.

Compare that to the *.DAT file - http://media.affymetrix.com/support/developer/powertools/changelog/gcos-agcc/dat.html. A DAT file is really only useful for most users in that it can be used to generate the feature summarized CEL file. And DAT files are not housed in public databases like NCBI's GEO anyway (even in-house, where I work, we don't keep *.DAT file for very long, we just archive the *.CEL files in our own database).

This is pretty much the norm nowadays for many genomic fluorescence scanning instruments - the actual image analysis is usually done in real time on the instrument, and only summarized data is output for further analyses by the end user. Especially in next-gen sequence instruments, the move to real-time image analysis (and thus, negating the need to store the images) drastically cuts down experimental run time and disc storage for output (e.g. ABI instruments, for example, use dedicated nVidia CUDA cards for real time fluorescence scan analysis, and only write scan summary data to disc).