Chipper: discovering transcription-factor targets from chromatin immunoprecipitation microarrays using variance stabilization
© Gibbons et al.; licensee BioMed Central Ltd. 2005
Received: 23 March 2005
Accepted: 30 September 2005
Published: 1 November 2005
Chromatin immunoprecipitation combined with microarray technology (Chip2) allows genome-wide determination of protein-DNA binding sites. The current standard method for analyzing Chip2 data requires additional control experiments that are subject to systematic error. We developed methods to assess significance using variance stabilization, learning error-model parameters without external control experiments. The method was validated experimentally, shows greater sensitivity than the current standard method, and incorporates false-discovery rate analysis. The corresponding software ('Chipper') is freely available. The method described here should help reveal an organism's transcription-regulatory 'wiring diagram'.
A major goal in understanding cellular behavior is to reveal the 'wiring' of transcriptional regulation, through which transcription factors (TFs) bind target-gene promoters to control gene expression. Promoter regions contain sequence elements - typically 5 to 12 nucleotides (nt) in length - at which TFs bind specifically. By enhancing/inhibiting transcription or recruiting complexes that remodel chromatin structure, TFs regulate expression of the genes whose promoters they bind. Chromatin immunoprecipitation (ChIP) is an experimental technique for identifying those regions of DNA bound by a particular protein, and is, therefore, a useful method for determining which genes have their promoters bound by a TF. In outline, the method consists of the following steps. The TF under study is crosslinked to DNA which is subsequently extracted and sheared into fragments approximately 400 nt long (1,000 nt resolution is usually sufficient to assign binding to the regulation of a specific gene, so it is rare to exceed this length ). The fragments are immunoprecipitated with an antibody specific to that TF (or to a peptide affinity tag fused to that TF), whereupon the crosslinks are reversed, the DNA precipitate amplified, and the intergenic regions (IGRs) containing the binding site(s) are determined by examining the relative abundance of each immunoprecipitated DNA fragment. The combination of ChIP with microarray technology is often called 'ChIP-chip'  and is referred to here as 'Chip2'. It has turned ChIP into a high-throughput technique for efficiently mapping gene regulatory networks [2–9].
Two-channel microarrays use hybridization to compare the abundance of specific nucleic acid sequences in one mixture to abundance of the same sequences in another control mixture. The choice of control mixture may greatly affect the outcome of the experiment. A typical choice is fragmented genomic DNA, which controls for the relative abundance and non-specific hybridization potential of genomic DNA fragments. Genomic DNA may be purified from 'whole-cell extract', which itself is sometimes used as a control. As some DNA fragments may be 'stickier' than others, a more stringent and laborious mock control (containing fragments recovered nonspecifically by immunoprecipitation (IP)) is sometimes performed, in which the TF does not have a fused affinity tag.
The change in abundance of a particular sequence between two mixtures is often measured in terms of 'fold-change' between the two channels (ratio) or, alternatively, the logarithm of fold-change (log-ratio). The IP channel serves as numerator, while the control is the denominator. The array surface between regions with spotted DNA is never completely 'dark', due to the combined effects of residual DNA fragments bound non-specifically to the array surface, and the experimentalist's control of the visual amplification ('gain') in the image analysis software. It is customary to subtract this 'background' from each spot because it reveals nothing about the protein-DNA binding. This subtraction raises the possibility, however, that the denominator could become negative or zero, in which case the log-ratio is not useful. Common strategies for handling zero or negative values are either to threshold or to discard data points altogether, neither of which is entirely satisfactory. A further, and perhaps more serious, problem is the practice of interpreting this fold-change as a measure of significance, when it provides no such statistical basis. Small random fluctuations in signals close to background, particularly in the denominator, are amplified, leading to spuriously high levels of 'fold-change' . In other words, we should reduce our confidence in a twofold change between signals that are each near the background noise, compared to a twofold change between strong signals. Because we are generally more interested in whether a region is specifically bound at all than we are in the degree of its binding (occupancy), there is a need for an accurate measure of confidence in each measurement.
A statistical approach for analysis of mRNA abundance microarrays has been developed in which a 'single-array' error model accounts for variation in the background level for each microarray, while a 'gene-specific' error model describes variation of a single gene across replicate arrays. These two complementary models can be combined to estimate the error in each log-ratio measurement . A variant of the single-array approach (in which there is gene-specific normalization) has been applied to transcription-factor binding site identification by means of Chip2 in yeast . Unfortunately, it requires one or more separate control experiments to determine error model parameters, in which identical nucleic acid mixtures are compared. This adds to the expense of the experiment; furthermore, error model parameters derived from a separate microarray are potential sources of systematic error, since quality can vary between microarrays.
Results and discussion
Here we describe a new approach for assessing statistical significance of TF-binding from Chip2 data. We illustrate our method using a Chip2 analysis of Sko1 (also known as Acr1), a TF of the basic leucine zipper (bZIP) family (CREB sub-family) that regulates the expression of osmotic stress inducible genes [11–13]. We also use independent confirmation experiments of individual IGRs to validate our method.
We distinguish two kinds of repeated experiment. When the same IGR is spotted onto an array in more than one location, we term these measurements 'duplicates,' and we consider them as two spatially separated parts of the same 'spot'. Though other approaches have been described , for simplicity we average duplicate signals before analyzing them, giving us a single value that is less susceptible to physical blemishes on the slide. When the same IGRs are spotted onto two or more distinct microarrays, we term them 'replicates.' We consider each replicate as an independent measurement of the binding affinity or 'occupancy' of the IGRs.
It is common to replicate genome-wide experiments several times, to improve confidence in the results, which may be degraded by array imperfections or by handling errors. Additional replicates can compensate for random error in individual measurements, and the typical number of replicates is likely to increase as the cost of microarrays falls . Sometimes the most significantly enhanced IGRs are those with low signal-to-noise ratio, yet applying log-ratios to such signals has the potential to introduce many false positives because minor variations in a small denominator value can have a large effect on a ratio. A single-array error model can account for this variation in calculating significance for each IGR. The log-ratios themselves are difficult to interpret, however, because two IGRs with the same log-ratio may differ in significance, and a greater log-ratio does not indicate increased significance. An alternative approach, the method of variance stabilization, was described by two groups [15, 16] and made available as part of the BioConductor project  in the package 'vsn' . It uses a regression algorithm that is robust to outliers to scale and offset each channel independently, in such a way that the variance between channels is independent of signal strength. The transformation of the signal y i in the ith channel (i = 1 for IP, or i = 2 for control) can be expressed as:
where α i and λ i represent the background and noise in the ith channel, respectively. Because ln(a) - ln(b) = ln(a/b), the difference between the two transformed channels (Δh ≡ h i - h2) is then a generalized log-ratio that is asymptotically equivalent to the log-ratio of the original channels when both are high (y i >> α i ), yet transforms smoothly to the difference between channels when both are low. This allows direct comparison between any two datapoints, even when they belong to opposite ends of the microarray's dynamic range. Two IGRs with the same Δh are equally significant, and greater Δh implies a more significantly bound IGR.
Deriving error model parameters internally
Binding of protein to DNA is a dynamic, stochastic process in equilibrium. While every TF is likely to be bound to every IGR at least some fraction of the time, our goal here is to perform binary classification of the IGRs. We therefore consider IGRs to fall into two categories: those that are specifically bound by the TF and those that are not. We wish to compute a p value that expresses our degree of surprise at seeing a particular Δh score for a given IGR, under the null hypothesis that the IGR is not bound. The 'vsn' package can be used to variance-stabilize each array separately, or all of them simultaneously; we used the former method. Having computed the inter-channel variance-stabilized difference (Δh) for each spot, we may plot a histogram of all scores from a chip. We expect that most regions are not bound. Therefore, the distribution of Δh scores should be largely determined by random binding and measurement errors . A smaller number of regions are bound, and those will tend to have positive scores, indicating higher occupancies in the IP channel than the whole-cell extract/mock control. Measurements in the negative portion of the Δh distribution should, therefore, be more completely dominated by unbound IGRs. By fitting a parametric curve to the region of the observed Δh distribution left of the mode, we obtain an estimate of the null distribution in the positive region of the Δh distribution. This is an essential feature of our method, because it allows us to estimate the distribution expected of unbound IGRs without performing an external control experiment in which an identical mixture is examined in both channels of a separate microarray. It is this null distribution that permits calculation of significance for each observed Δh value. The symmetric nature of the null distribution is an assumption of our model, and is based on our own experience and that of others .
Experimental verification of our dataset and evaluation of pvalue accuracy
We independently validated 35 target genes spread widely across the top 350 in our list using targeted ChIP analysis. Considering only the 35 targets for which follow-up testing was performed, ranking of IGRs by the p values of Lee et al.  (see Additional data file 4) shows an ability similar to our method ('Chipper') at placing true positives above false positives. When considering all IGRs, however, there is little correlation between rank by our method and rank by the Lee et al. approach. In other words, top-ranking targets by one method are not top-ranking by the other. Thus, although our validation experiments are consistent with Chipper achieving the same sensitivity at a lower false-positive rate, it is also possible that the two methods are each adept at identifying different subsets of targets. The discrepancy may be due to some systematic error in determination of the parameters of the error model. As the error model parameters are not provided explicitly with their data, we could not investigate this possibility further. Inaccurate determination of error-model parameters can lead to unjustified confidence in differences based on noisy measurements. Therefore, in the task of ranking IGRs by the likelihood of being TF-bound, Chipper is on par and complementary to the Lee et al. approach and may outperform it. Furthermore, the Chipper algorithm uses an internally determined error model and thus is not subject to systematic errors that may arise via the separate control experiments required of the methods in Lee et al. . Below we show that Chipper allows increased sensitivity at a given significance threshold.
Divergently transcribed genes, grouped in pairs of which at least one is a target of Sko1, according to a targeted ChIP assay
False discovery rate analysis
Validation with publicly available datasets
In practice, however, it is common to consider only those IGRs passing a standard threshold of significance (p < 10-3 in Lee et al.  and Harbison et al. ). Therefore, we evaluated the same data, but rather than focusing on simple ranking ability, we examined the p value of each call (results for Ino4 shown in Figure 4b). We constructed the graph by choosing a significance threshold (α) and asking what fraction of the known true positives exceed the threshold (that is, have p values less than α). At α = 1, any algorithm will have perfect sensitivity because it calls all IGRs significant; this comes at the cost of specificity, as it is unable to distinguish between true and false positives. The p values reported by Lee et al.  are shown in green, those by our method are shown in black. The vertical dotted line indicates a threshold α7 = 10-3 at which we would expect approximately 7 out of 7,200 intergenic regions to achieve significant scores purely by chance, even if none were bound by the TF. The vertical dashed line indicates the threshold α1 = 1.6 × 10-4, which we expect to be exceeded by chance for only one out of 7,200 IGRs. The unshaded area to the right of α1 indicates the region in which fewer than one IGR would be expected to exceed the threshold by chance. The higher an algorithm's sensitivity in this region (that is, the more true positives it puts here), the better. As we decrease the threshold, the sensitivity decreases slowly at first, for both methods. For the p values of Lee et al. , there is then a rapid reduction in sensitivity. At an α threshold such that only one false positive is expected, our method can recover more than half the known targets while Lee et al.  find none.
In Figure 4c, we show an ROC curve for the transcription factor Sko1, for which nine targets are annotated in the YPD. The error model of Lee et al.  ranks the targets slightly better than our method of average z scores. Yet, as shown in Figure 4d, for any given significance threshold, our algorithm returns more of those targets. Ino4 showed the most striking improvement in sensitivity (Figure 4b) for all TFs examined. However, for each of the eight TFs we examined (Figure 4 and Additional data file 5) our method called an equal or greater number of targets significant at the level of α1 than did the method of Lee et al. . Thus, for all TFs examined, our method yields sensitivity either markedly better than or similar to that of the de facto standard method.
We have developed a method for analyzing results from chromatin-immunoprecipitation/microarray (Chip2) experiments that computes p values without needing a separate control for developing a model of measurement error. The method proposed here successfully combines multiple replicates (separate arrays) and duplicates (same array) to produce a single overall p value for each IGR. By using variance stabilization rather than log ratios, we eliminate the need to threshold low-signal spots obtaining an alternative measure, Δh, which interpolates between a difference and a log-ratio and is monotonically related to significance. In addition, by averaging the resulting z score over replicates, an IGR that scores highly in a single replicate, but has no usable data in other replicates, may score well in the overall rankings. This is desirable in hypothesis generation: the algorithm should not be conservative, rather it should be sensitive and provide accurate p values by which the false positive rate can be judged. The p values produced by our algorithm behave as one would expect p values to: a broadly uniform distribution over the full range, but with enrichment near p = 0. Experimentalists can use the q values computed from these p values to generate a short list that is customized to their tolerance for false discoveries. We have evaluated our algorithm using the transcription factor Sko1 by performing targeted ChIP on 35 selected genes. Additionally, we have compared performance of our algorithm with that of a previous error model , using data from a public database of transcription-factor targets [28, 29]. Generally, discrimination of true positives, as measured by ROC curves, is comparable for both methods. However, our method returns targets with more significant p values. We find that the observed false-discovery rate on these putative targets generally tracks that predicted by the q values, therefore validating the accuracy of the p values and q values produced by our method. To parameterize error models, the method presented here requires no external control microarray experiments (which may introduce systematic error), giving it a distinct advantage over others in current use. Software implementing the algorithm is available either in web-based form for online use, or for download by non-commercial users, from our website .
Materials and methods
Chip2 analysis on Sko1 was performed using three microarrays, each with duplicate spots. Genomic DNA was used as a negative control. We used targeted ChIP experiments on 35 putative targets of Sko1 to validate how well our algorithm finds TF binding sites. We selected targets distributed throughout the top-ranking 350 IGRs. Primers were specifically designed for each IGR, and each region was assayed three times both with and without the hemagglutinin (HA) epitope tag, and the results averaged. The POL1 open reading frame (ORF) and an ORF-free region were used as negative controls, since Sko1 is not expected to bind there. Each IGR was scored according to the ratio of its IP efficiency with the HA epitope tag compared to that of POL1 ORF (non-specific control). Based on prior experience, we chose a threshold of 2.0, above which we considered Sko1 to have bound to the IGR, and below which we considered it not to have bound. By this criterion, we found 21 bound IGRs, with the remaining 7 tested IGRs not bound. (The number of IGRs tested is less than the number of target genes because some IGRs are associated with more than one gene.) Of those scoring >2.0, we found that six (ICY1, HOR7, YPR127W, DPM1, POS5, and RSN1) also scored highly (above 2.0) without the tag, indicating that they bind non-specifically. In fact, only POS5 scored in the top 100 by our method. Further details on Chip2 analysis of Sko1 and validation experiments are published elsewhere in the context of a focused study of Sko1 . The complete dataset is available from the Gene Expression Omnibus (GEO)  under series accession number GSE3335.
Additional data files
The following additional data are available with the online version of this paper. Additional data file 1 is a tab-delimited file containing the results of our analysis for all IGRs studied in our experiments. Additional data file 2 contains a detailed description of the comparison between the targets of Sko1 identified by Chipper when applied both to the data presented here and to other Chip2 data , and previously published p values using a single-array error model . Additional data files 3 and 4 are figures illustrating these comparisons. Additional data file 5 is a figure comparing the two methods as applied to results from six additional transcription factors. Additional data file 6 lists the IGRs identified as targets .
We thank J Geisberg, M Damelin, P Silver, Z Moqtaderi and J Wade for helpful discussions, and J Geisberg and J Casolari for 'beta-testing' the website and algorithm. F.D.G. and F.P.R. were supported in part by Funds for Discovery provided by John Taplin and by an institutional grant from the HHMI Biomedical Research Support Program for Medical Schools. M.P., F.D.G., and K.S. were supported by NIH/NIGMS grants GM30186, GM53720, and NIH/NHGRI grant HG003147. M.P. was supported by an EMBO Long Term Fellowship and the 'Ramón y Cajal' program of the Spanish Ministry of Science.
- Buck MJ, Lieb JD: ChIP-chip: considerations for the design, analysis, and application of genome-wide chromatin immunoprecipitation experiments. Genomics. 2004, 83: 349-360. 10.1016/j.ygeno.2003.11.004.PubMedView ArticleGoogle Scholar
- Lee TI, Rinaldi NJ, Robert F, Odom DT, Bar-Joseph Z, Gerber GK, Hannett NM, Harbison CT, Thompson CM, Simon I, et al: Transcriptional regulatory networks in Saccharomyces cerevisiae. Science. 2002, 298: 799-804. 10.1126/science.1075090.PubMedView ArticleGoogle Scholar
- Iyer VR, Horak CE, Scafe CE, Botstein D, Snyder M, Brown PO: Genomic binding sites of the yeast cell-cycle transcription factors SBF and MBF. Nature. 2001, 409: 533-538. 10.1038/35054095.PubMedView ArticleGoogle Scholar
- Lieb JD, Liu X, Botstein D, Brown PO: Promoter-specific binding of Rap1 revealed by genome-wide maps of protein-DNA association. Nat Genet. 2001, 28: 327-334. 10.1038/ng569.PubMedView ArticleGoogle Scholar
- Ren B, Robert F, Wyrick JJ, Aparicio O, Jennings EG, Simon I, Zeitlinger J, Schreiber J, Hannett N, Kanin E, et al: Genome-wide location and function of DNA binding proteins. Science. 2000, 290: 2306-2309. 10.1126/science.290.5500.2306.PubMedView ArticleGoogle Scholar
- Pugh BF, Gilmour DS: Genome-wide analysis of protein-DNA interactions in living cells. Genome Biol. 2001, 2: reviews1013.1-1013.3. 10.1186/gb-2001-2-4-reviews1013.View ArticleGoogle Scholar
- Ng HH, Robert F, Young RA, Struhl K: Genome-wide location and regulated recruitment of the RSC nucleosome-remodeling complex. Genes Dev. 2002, 16: 806-819. 10.1101/gad.978902.PubMedPubMed CentralView ArticleGoogle Scholar
- Harbison CT, Gordon DB, Lee TI, Rinaldi NJ, Macisaac KD, Danford TW, Hannett NM, Tagne JB, Reynolds DB, Yoo J, et al: Transcriptional regulatory code of a eukaryotic genome. Nature. 2004, 431: 99-104. 10.1038/nature02800.PubMedPubMed CentralView ArticleGoogle Scholar
- Cawley S, Bekiranov S, Ng HH, Kapranov P, Sekinger EA, Kampa D, Piccolboni A, Sementchenko V, Cheng J, Williams AJ, et al: Unbiased mapping of transcription factor binding sites along human chromosomes 21 and 22 points to widespread regulation of noncoding RNAs. Cell. 2004, 116: 499-509. 10.1016/S0092-8674(04)00127-8.PubMedView ArticleGoogle Scholar
- Hughes TR, Marton MJ, Jones AR, Roberts CJ, Stoughton R, Armour CD, Bennett HA, Coffey E, Dai H, He YD, et al: Functional discovery via a compendium of expression profiles. Cell. 2000, 102: 109-126. 10.1016/S0092-8674(00)00015-5.PubMedView ArticleGoogle Scholar
- Nehlin JO, Carlberg M, Ronne H: Yeast SKO1 gene encodes a bZIP protein that binds to the CRE motif and acts as a repressor of transcription. Nucleic Acids Res. 1992, 20: 5271-5278.PubMedPubMed CentralView ArticleGoogle Scholar
- Proft M, Serrano R: Repressors and upstream repressing sequences of the stress-regulated ENA1 gene in Saccharomyces cerevisiae: bZIP protein Sko1p confers HOG-dependent osmotic regulation. Mol Cell Biol. 1999, 19: 537-546.PubMedPubMed CentralView ArticleGoogle Scholar
- Vincent AC, Struhl K: ACR1, a yeast ATF/CREB repressor. Mol Cell Biol. 1992, 12: 5394-5405.PubMedPubMed CentralView ArticleGoogle Scholar
- Smyth GK, Michaud J, Scott H: Use of within-array replicate spots for assessing differential expression in microarray experiments. Bioinformatics. 2005, 21: 2067-2075. 10.1093/bioinformatics/bti270.PubMedView ArticleGoogle Scholar
- Huber W, von Heydebreck A, Sültmann H, Poustka A, Vingron M: Variance stabilization applied to microarray data calibration and to the quantification of differential expression. Bioinformatics. 2002, 18 Suppl 1: S96-S104.PubMedView ArticleGoogle Scholar
- Durbin BP, Harin JS, Hawkins DM, Rocke DM: A variance-stabilizing transformation for gene-expression microarray data. Bioinformatics. 2002, 18 Suppl 1: S105-S110.PubMedView ArticleGoogle Scholar
- Gentleman RC, Carey VJ, Bates DM, Bolstad B, Dettling M, Dudoit S, Ellis B, Gautier L, Ge Y, Gentry J, et al: Bioconductor: open software development for computational biology and bioinformatics. Genome Biol. 2004, 5: R80-10.1186/gb-2004-5-10-r80.PubMedPubMed CentralView ArticleGoogle Scholar
- Rocke DM, Durbin B: A model for measurement error for gene expression arrays. J Comput Biol. 2001, 8: 557-569. 10.1089/106652701753307485.PubMedView ArticleGoogle Scholar
- Huber W, von Heydebreck A, Sueltmann H, Poustka A, Vingron M: Parameter estimation for the calibration and variance stabilization of microarray data. Stat Appl Genet Mol Biol. 2003, 2: 3.1-3.22.Google Scholar
- Dennis JE, Schnabel RB: Numerical Methods for Unconstrained Optimization and Nonlinear Equations. 1983, Englewood Cliffs, NJ: Prentice-HallGoogle Scholar
- Press WH, Flannery BP, Teukolsky SA, Vetterling WT: Numerical Recipes. 1986, Cambridge, UK: Cambridge University Press, 1Google Scholar
- Proft M, Gibbons FD, Copeland M, Roth FP, Struhl K: Genomewide identification of Sko1 target promoters reveals a regulatory network that operates in response to osmotic stress in Saccharomyces cerevisiae. Eukaryotic Cell. 2005, 4: 1343-1352. 10.1128/EC.4.8.1343-1352.2005.PubMedPubMed CentralView ArticleGoogle Scholar
- Storey JD: The positive false discovery rate: a Bayesian interpretation and the q-value. Ann Statistics. 2003, 31: 2013-2035. 10.1214/aos/1074290335.View ArticleGoogle Scholar
- Sokal RR, Rohlf FJ: Biometry: The Principles and Practice of Statistics in Biological Research. 1995, New York: WH Freeman & Company, 3Google Scholar
- Benjamini Y, Hochberg Y: Controlling the false discovery rate: a practical and powerful approach to multiple testing. J R Stat Soc Ser B. 1995, 57: 289-300.Google Scholar
- Storey JD, Tibshirani R: Statistical significance for genomewide studies. Proc Natl Acad Sci USA. 2003, 100: 9440-9445. 10.1073/pnas.1530509100.PubMedPubMed CentralView ArticleGoogle Scholar
- Payne WE, Garrels JI: Yeast Protein Database (YPD): a database for the complete proteome of Saccharomyces cerevisiae. Nucleic Acids Res. 1997, 25: 57-62. 10.1093/nar/25.1.57.PubMedPubMed CentralView ArticleGoogle Scholar
- Costanzo MC, Hogan JD, Cusick ME, Davis BP, Fancher AM, Hodges PE, Kondu P, Lengieza C, Lew-Smith JE, Lingner C, et al: The yeast proteome database (YPD) and Caenorhabditis elegans proteome database (WormPD): comprehensive resources for the organization and comparison of model organism protein information. Nucleic Acids Res. 2000, 28: 73-76. 10.1093/nar/28.1.73.PubMedPubMed CentralView ArticleGoogle Scholar
- Csank C, Costanzo MC, Hirschman J, Hodges P, Kranz JE, Mangan M, O'Neill K, Robertson LS, Skrzypek MS, Brooks J, et al: Three yeast proteome databases: YPD, PombePD, and CalPD (MycoPathPD). Methods Enzymol. 2002, 350: 347-373.PubMedView ArticleGoogle Scholar
- Chipper. [http://llama.med.harvard.edu/Software.html]
- Barrett T, Suzek TO, Troup DB, Wilhite SE, Ngau WC, Ledoux P, Rudnev D, Lash AE, Fujibuchi W, Edgar R: NCBI GEO: mining millions of expression profiles - database and tools. Nucleic Acids Res. 2005, 33 Database issue: D562-D566.Google Scholar
- QVALUE: The Manual. Version 1.0. [http://faculty.washington.edu/~jstorey/qvalue/manual.pdf]
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.