IntroductionChromatin immunoprecipitation coupled to tiling microarray analysis (ChIP-on-chip) is used to measure genome-wide the DNA binding sites of a protein of interest. In ChIP-on-chip, proteins are covalently cross-linked to the DNA by formaldehyde, cells are lysed, the chromatin is immunoprecipitated with an antibody to the protein of interest and the fragmented DNA that is directly or indirectly bound to the protein is analyzed with tiling arrays. For this purpose, the fragmented DNA is fluorescently labeled and hybridized to the tiling array, which consists of millions of short (25 to 60 nucleotides long) probes that cover the genome at a constant spacing (4 to 100s of nucleotides), like tiles covering a roof. The data generated by one experiment consists of an intensity value for each DNA probe. These values measure the relative quantity of DNA at the probe\'s genomic position in the immunoprecipitated material.This guideline describes the first steps in the data analysis for ChIP-on-chip measurements in a bare-bones fashion. The steps comprise the quality control of the obtained data and normalizations to render the data comparable between different arrays, to correct for saturation effects, and to obtain enrichment and occupancy values. Peak calling and other downstream procedures are not part of this exposé. For each step, we give detailed recommendations, warn about problematic but often popular procedures, and occasionally suggest improved versions of standard procedures.Although we have gained experience in ChIP chip data analysis mainly in yeast, this protocol tries to give a general guideline that should be applicable to other species and to both single and two-color arrays. Most advocated procedures can be carried out within the R environment for statistical data analysis, using packages from the Bioconductor project (see, e.g., Toedling and Huber, 2008). A \"Bioconductor package Starr for Affymetrix platforms supporting the analysis steps described here is available and will be described in an upcoming protocol (Zacher B. and Tresch A., 2010).Back to topProcedureExperimental DesignNumber of ReplicatesIt is advisable to measure at least two biological replicates of all factors or conditions, for two reasons. First and foremost, corrupted measurements will normally be easily identifiable by comparing replicates. Even if both of two measurements are corrupted, corruption tends to be erratic and will normally not result in the expected high correlation between replicates (step 2 below). Second, averaging over N replicates reduces the standard deviation of unsystematic noise (i.e. the random scattering of measured values) by a factor vN. This means a factor of 1.4 reduction for two replicate measurements, 1.7 for three, and 2.0 for four replicates. The pay-off is thus particularly high for the second replicate and falls of slowly for higher N.Mock IP or Genomic Input?In any ChIP-on-chip experiment, it is important to correct for sequence- and genomic region-specific biases in the efficiency of the various biochemical and biophysical steps of the ChIP-on-chip protocol. Crosslinking of the DNA to proteins, fragmentation of the chromatin, immunoprecipitation, PCR amplification, and hybridization to the array all have strong biases that must be corrected for. This is done by measuring a reference signal and dividing the true signal intensities obtained from the immunoprecipitated protein by the reference intensities. This is in our experience by far the most important step in data normalization. A frequent point of contention is whether it is better to use the genomic input fraction or a mock immunoprecipitation (mock IP) for this normalization. (The mock IP is performed with the wild-type strain if the true signal is obtained by immunoprecipitating the protein of interest with a protein tag. It is done using an unspecific antibody if the protein of interest is purified with a specific antibody.)Our experience is that normalization with the mock IP is preferable if a good signal can be obtained from the mock IP hybridization, i.e., if the signal-to-noise ratio of the mock IP is not much inferior to that of the true IP. We have seen higher noise levels and occasional artefacts when using normalization with the genomic input. One reason might be that normalization with the genomic input does not correct for sequence- and region-dependent biases of unspecific binding to the antibody. If signal-to-noise of the mock is a problem, or if the chromatin preparation is deemed hard to reproduce in different strains, then normalization with matched genomic input is advisable. An obvious drawback for single-color arrays is, however, that for each measurement, a matched genomic input is hybridized, necessitating almost twice the number of arrays as if all measurements are normalized with a single mock IP. We have found that genomic input measurements can be highly reproducible for different experiments and chromatin preparations. In this case it may be justified to use only a few representative input measurements for data normalization and to forego the costly measurement of a matched genomic input sample for each IP (see comment 1).A reasonable procedure is to measure both input IP and mock IP for representative factors and to compare the signal-to-noise ratio for both normalizations. Another possibility for two-color arrays is to normalize with both input and mock IP:log enrichment = log [ (signal1/input1) / (mock2/input2) ] = log(signal1/input1) - log(mock2/input2)In this case, a dye exchange should be performed (next paragraph). Dye ExchangeWhen using two-color arrays, such as those from NimbleGen or Agilent, the dye-exchange procedure is very much recommended (e.g. Do and Choi, 2006). This means the dyes Cy3 and Cy5, with which signal (true IP) and reference fractions are labeled, should be switched between the two (or four) biological replicates, such that the signal fraction is labeled with Cy5 in half the measurements and with Cy3 in the other half. Dye exchange corrects for dye-dependent saturation effects in a much better way than downstream data normalization could ever achieve (steps 5 and 5a below).[1][2][3]下一页