Can Zipf's law be adapted to normalize microarrays?
© Lu et al; licensee BioMed Central Ltd. 2005
Received: 30 August 2004
Accepted: 23 February 2005
Published: 23 February 2005
Normalization is the process of removing non-biological sources of variation between array experiments. Recent investigations of data in gene expression databases for varying organisms and tissues have shown that the majority of expressed genes exhibit a power-law distribution with an exponent close to -1 (i.e. obey Zipf's law). Based on the observation that our single channel and two channel microarray data sets also followed a power-law distribution, we were motivated to develop a normalization method based on this law, and examine how it compares with existing published techniques. A computationally simple and intuitively appealing technique based on this observation is presented.
Using pairwise comparisons using MA plots (log ratio vs. log intensity), we compared this novel method to previously published normalization techniques, namely global normalization to the mean, the quantile method, and a variation on the loess normalization method designed specifically for boutique microarrays. Results indicated that, for single channel microarrays, the quantile method was superior with regard to eliminating intensity-dependent effects (banana curves), but Zipf's law normalization does minimize this effect by rotating the data distribution such that the maximal number of data points lie on the zero of the log ratio axis. For two channel boutique microarrays, the Zipf's law normalizations performed as well as, or better than existing techniques.
Zipf's law normalization is a useful tool where the Quantile method cannot be applied, as is the case with microarrays containing functionally specific gene sets (boutique arrays).
DNA microarrays have become a widely used biotechnology for assessing expression levels of tens of thousands of genes simultaneously in a single experiment [1, 2]. Whether microarrays are being used for global tissue profiling or for differential expression studies, data normalization is an essential preliminary step before statistical analysis methods can be applied. The purpose of all normalization techniques is to transform the data to eliminate sources of variability stemming from experimental conditions, leaving only biologically relevant differences in gene expression for subsequent analysis. Normalization can be divided into two stages, intra-array normalization and inter-array normalization. Intra-array normalization deals with variability within a single array caused by factors such as differences in print-tip characteristics, channel differences in two-dye systems, and spatial heterogeneity across the array surface [3–5] and should be carried out using accepted methods before inter-array normalization is applied. This paper assumes intra-array normalization has been performed and presents an inter-array normalization method for comparison of gene intensity levels between multiple microarrays to deal with variation caused by such factors as differences in RNA isolation efficiency, labeling efficiency, hybridization conditions, exposure times, and detection efficiencies.
It is now clear that simple inter-array normalization techniques, such as simple scaling to housekeeping genes or normalizing to a global mean, are not adequate for microarray data . Housekeeping genes have been found to be more susceptible to modulation than previously thought . Along with others , this paper underscores the potentially serious drawbacks of the global mean and other such methods. Recent literature has thus provided a plethora of more sophisticated normalization and analysis techniques as researchers struggle to cope with the task of microarray data analysis, some of which include maximum likelihood analysis , centralization , principal component analysis , analysis of variance  and Bayesian network analysis .
Analysis of publicly available large-scale SAGE gene expression data sets [11, 12] and an intra-phyletic survey of genome wide Affymetrix microarray experiments  have indicated that the large majority of expressed genes exhibited power-law distributions, while some microarray expression data exhibit a more log-normal distribution . Our normalization procedure was inspired by the observation that the intensities measured on our microarray system also followed a power law distribution and can therefore be described by a simple mathematical model. Zipf's law  is a power law function that states that the magnitude of an intensity measurement (y) is inversely proportional to the rank (r) of that data point in the data set,
y∝r c (1)
where c is a coefficient close to -1. Our microarray data can be classified as a generalized form of Zipf's law because the coefficient (c) is not always close to -1 and, in fact, varies between individual microarrays, making simple linear normalization procedures, such as global normalization to the same mean, inappropriate. However, the normalization procedure proposed here demonstrates that by taking Zipf's law into account, it is possible to apply a simple intra-array normalization procedure such that all filters have the same coefficient c and proportionality.
We demonstrate the Zipf's law based normalization technique on microarray data sets representing both single channel and two channel technologies. In the single channel category, we produced two radio-labeled, nylon membrane based cDNA data sets, one commercial and one generated "in-house". Both systems contain a selection of genes chosen without regard to functional or pathway considerations, which make them especially appropriate for normalization using Zipf's law. These data sets were also normalized to a global mean (the mean of all microarrays) , and the quantile normalization method . In addition we produced a two channel, fluorescently labeled, glass slide, oligo-based microarray data set generated 'in-house'. This microarray can be classified as a 'boutique' microarray because it consists of a selection of genes involved in apoptosis. This data set was normalized with a variant of the Zipf's law normalization method that uses a subset of the distribution as a proxy for normalizing the entire microarray. A comparison was then conducted against a variant of the loess normalization method that uses an a priori selection of 'housekeeping' genes as a proxy for normalization.
The finding that our microarray data distributions conform to a power law distribution agrees with predictions based on genome wide gene expression studies [11–13], however Hoyle, et. al.  observed that microarray distributions were log normally distributed with possible power law tails. To investigate this discrepancy, and to verify that our normalization technique could be useful in the normalization of data sets from other microarray systems, we also surveyed publicly available data sets from the NCBI Gene Expression Omnibus .
The two assumptions upon which the normalization method are based are the same as those used in other normalization methods [5, 6], namely that in comparisons between similar tissues or cell lines under different experimental conditions i) most genes are not, or only moderately, regulated, and ii) approximately equal numbers of genes are up regulated as down regulated. Systems which conform to these two assumptions will be referred to as 'well-behaved' in this paper. While these assumptions probably hold for microarrays derived from a diverse sampling of genes, for example an EST library survey, they may not hold for microarrays containing genes specifically selected based on function or pathway (so called 'boutique' microarrays) as it is likely that most genes will be affected by the experimental treatments. One way to circumvent the restrictions resulting from these assumptions is to use a subset of data, or proxy, from the boutique array data set which fulfils the 'well-behaved' criteria. In developing a boutique microarray normalization technique, Wilson et. al.  have devised a method for selecting a subset of genes within a microarray data set that have low variation between arrays and are well representative of the spectrum of intensities measured on the microarray. They term this a priori selected subset 'housekeeping' genes, however it should not be confused with the a posteriori set of genes typically envisioned when the term is used. Another possible proxy that could meet the 'well-behaved' criteria are control spots which are included on the microarray during it's manufacture. We tested our normalization method on data from a two channel boutique microarray experiment using two types of control spots as proxies for normalization (Positive and negative internal controls, and housekeeping genes). The Zipf's law normalization methods were then compared with the variant of the loess method developed by Wilson et. al.  using housekeeping genes.
Verifying Zipf's Law
Data set comparison
Number of Data Points
Number of Expts
Human Unigene RZPD 1
Clontech Atlas Rat cDNA Expression
Clontech Atlas Human 1.2 (I & II)
Clontech Atlas Mouse 1.2
Clontech Atlas Human Cancer 1.2
NlaIII: Rattus norvegicus
NlaIII: Homo sapiens
5 × 2
Caltech 16K cDNA mouse
Stanford Human Unigene
Affymetrix GeneChip Rat Genome
Affymetrix GeneChip Human Genome
Normalization results – single channel microarrays
A comparison of the Zipf's law normalization method to the simple method of setting all arrays to a global mean (the mean of all microarrays) and to the quantile method was conducted on the single channel microarray data sets. Five human Unigene microarrays from the panel of thirty-two microarrays used in the sigmoidal colon experiments were selected to represent the greatest variability in pre-normalized data observed in the experiment (Figure 1b). Normalization to a global mean (Figure 1c) yielded data sets that displayed a higher variability in the coefficient c of the Zipf's power function (formula 1) than that observed after normalization by the Zipf's law method (Figure 1e) or the quantile method (Figure 1d). The Zipf's method showed the lowest variation in the Zipf's exponent and had the lowest spread of the data around the ln(rank) vs. ln(intensity) line. Results of an identical loge intensity versus loge rank plot comparison in Clontech rat microarrays showed little difference between the quantile and Zipf's methods [see Additional file 1]. However it should be mentioned that this method of data plotting provides one view of the data which is especially favorable to the Zipf's law normalization method. Next we examine the results of the MA-plots, a technique that is especially favorable to the quantile normalization method.
Pairwise Combinations (within array)
Clontech Atlas Rat cDNA Expression
Human Unigene RZPD 1
Mouse Apoptosis Subset
Normalization Results – Two Channel Boutique Microarray
It should be noted that much of the variation in c coefficients under the various normalization regimes is due to one channel (Cy3) on one microarray which had low median intensity and high variance due to low labelling efficiency (depicted in black in Figure 4). When normalized with the loess techniques (Figure 4c and 4f) the second channel (Cy5) on this array is adjusted to have a similar median intensity and variance, possibly skewing the results in favour of the Zipf's normalization techniques. To make the normalization method comparison unbiased, we eliminated this array from the analysis [see Additional file 3]. The Zipf's normalization based on internal controls (a) showed the lowest variation in c coefficients, the methods based on selected housekeeping genes (b, c) performed approximately equally well. Here again, raw (d), quantile normalized (e), and loess normalized (f) plots are provided for reference only.
Finally, to quantify the differences between distributions after normalization, pairwise Kolmogorov-Smirnov values were computed for both the complete boutique array data set (Table 2c) and after eliminating the array which contained a low median intensity and high variance due to low labelling efficiency (Table 2d). In addition to computing the Kolmogorov-Smirnov values for all possible between-array pairwise combinations, we also summarized just the within-array pairwise comparisons (in parenthesis in Table 2). Of the normalization methods which can be applied to boutique microarrays, the Zipf's method using internal controls produced the most similar data distributions when all possible between-array comparisons are taken into consideration. When only within-array comparisons are considered, the Zipf's method using internal controls was better after the low labelling efficiency array was eliminated. The Zipf's method using selected housekeeping genes did not perform as well as the other methods in within-array comparisons, and was the middle performer when all possible between-array comparisons were computed. Kolmogorov-Smirnov values were computed from the global mean, Zipf's general, quantile, and loess normalization methods and are provided for reference only.
Microarray platform comparison
In a survey of seventeen microarray data sets, Hoyle et. al.  reported that microarray data follow a log-normal distribution with power-law tails. The three data sets presented in this paper exhibited distinct power-law distributions (Table 1, data sets A, B and H). To ascertain whether the data sets we used were unusual, we surveyed nine additional data sets (Table 1, data sets C-G, I-K) to determine their conformity to Zipf's law and the log-normal distribution respectively. The microarray data sets fell into two broad categories, power law distributions (Figure 6, data sets A-E) and log normal distributions (Figure 6, data sets I-K). Of the six power law data sets, two (B and C) clearly followed Zipf's law distributions. The remaining four (data sets A, D, E, and H), while still power-law distributed, showing noticeable deviations from the distribution at the lower rank (higher intensity) portion of the distribution. Of the platforms that where recognizably log normal in distribution, two fluorescent dye labeled, oligo-based Affymetrics platforms (data sets K and L) followed the distribution most closely and two dye labeled, cDNA systems (data sets I and J) were perceptibly log normal. The two SAGE experiments (data sets F and G) which were included for comparison purposes, exhibited Zipf's law distributions. Coefficients of determination (r2) of the log mean intensity vs. log rank are a measure of conformation to a power-law distribution and ranged from 0.9968 to 0.7773 for microarray data sets, 0.9982 and 0.9978 for the SAGE experiments (Table 1).
Zipf's law is based on observations made by linguist George Kingsley Zipf that the frequency of word occurrences in natural languages is proportional to the negative power of the rank order of the word. Beside the original findings in natural languages , Zipf's law has been found to apply to a plethora of natural phenomena, from the populations of cities to the impact factors of scientific journals as well as a variety of biological data, of which a review made available by Wentian Li  is an excellent online resource. It is important to point out, that being a phenomenological principle, Zipf's law does not imply that there is a universal underlying physical process at work. However, in much the same way that the Gaussian-Normal distribution occurs naturally in data and can be used to statistically test or otherwise manipulate the data, the fact that microarray data conforms to Zipf's law can be adapted for the purpose of microarray normalization.
Zipf's law is a power law function that states that the magnitude of an intensity measurement is inversely proportional to the rank of that data point in the data set, where c is a coefficient close to -1. Ranking is a method common in statistics, which has previously been used to analyze microarray data. Hoyle el. al.  used ranking as a method for evaluating microarray data and proposed the use of several statistics including χ2 to quantify the agreement of the distribution to Benford's Law , and σ2 as a quality control measure to detect such factors as low signal to background ratio, or mRNA probes extracted from mixed cell types. Ranking also figured prominently in the evaluation of a survey of inter-array normalization methods  where the statistics 'absolute rank deviation' and 'relative rank deviation' were used to select the method that produces the most 'well-normalized' data. The normalization procedure described in this paper is the first to combine these two ideas, namely that ranking can be used to judge the effectiveness of a normalization method, and that microarray data conforms to Zipf's law. We evolved these ideas into a novel and easily applicable normalization method and compared this method with existing methods to eliminate non-biological variation from microarray data sets.
In order to implement an appropriate data normalization technique, it is important to know the distribution of a given data set. Several publications have examined the data distributions that typically result from microarray experiments. In a survey of seventeen microarray data sets, sixteen of which were fluorescent dye labeled, Hoyle et. al.  reported that microarray data were found to have a log normal distributions with power law tails. More recent publications have reported that the abundance of expressed genes exhibit power-law distributions [11, 13, 24]. Results from our own data sets and a subsequent survey of publicly available data sets from both radioactively and fluorescently labeled platforms suggest that both types of distributions can be manifested in microarray data.
Comparisons between the Zipf's law and quantile normalization methods using MA plots showed that the quantile method effectively removes intensity dependant effects, sometimes referred to as 'banana' curves, from microarray data sets, while the Zipf's law method has no effect on the curved nature of the intensity dependent effect. This is not altogether unexpected as the quantile method was specifically designed to remove such effects. While the Zipf's method does not remove the curve from the intensity dependent effect, it does minimize negative consequences by rotating the data distribution such that the maximal number of data points lie on the zero of the log ratio axis. In this respect, the Zipf's law normalization technique can be considered inferior to the quantile method, however, it may still be a useful tool where the quantile method cannot be applied.
One such case, in which quantile normalization is inappropriate, is with so called 'boutique' microarrays where the genes spotted on the array represent a selected set of genes, for example from a specific pathway or those involved with a particular biological process or disease state. In such systems, most genes are expected to be differentially regulated when control and experimental samples are compared and the expected data distribution of control samples may be significantly different than that of experimental samples (in mean intensity for example). The quantile normalization method would effectively remove this difference by replacing the data distribution of each microarray with the mean distribution of all arrays. In contrast, the principle of normalization according to Zipf's law can also apply to arrays of this type if a group of control spots are included on the microarray. These control spots could be an external reference probe which hybridises to a concentration gradient of matching spots on the array, or internal positive (highly expressed genes) and negative (spotting buffer) control spots on the microarray, or an a priori selected set of housekeeping genes using a method such as that described by Wilson et. al.  or Schadt et. al. . A linear model can be fitted to the control spots alone, and the normalization procedure can then be applied using the control spots as a proxy for the entire data distribution. The critical assumption in using control spots in normalization is establishing their relationship to the experimental spots.
The results of our comparison between methods which are designed to normalize boutique microarray data show that Zipf's law normalization using internal control spots results in a relatively well normalized data set when compared to Zipf's law normalization using selected housekeeping genes and the modified loess method using selected housekeeping genes. In addition, the Zipf's law method produced data distributions which are more similar between arrays allowing for between-array comparisons which are advantageous in terms of both cost, because of the reduced number of microarrays that need to be run, and, statistical power, by allowing for greater numbers (n), experimental design permitting.
In summary, we examined the applicability of using Zipf's law as the basis for a novel normalization technique, which is applicable to both one channel microarray data and two channel microarrays. This method is shown to out-perform such methods as global normalization to the mean but would appear to be inferior to quantile normalization. The quantile method was superior to Zipf's law in removing intensity dependent effects commonly seen in microarray data. While the latter method cannot be applied to boutique arrays, we show that the Zipf's normalization method used with internal positive and negative controls or with selected housekeeping genes normalizes boutique arrays as well as currently existing methods. Additionally, data normalized with the Zipf's method using internal control spots seems more amenable to between-array gene intensity comparisons when compared to other methods.
Data set A (Table 1) was generated using a global genome-wide cDNA clone set (Human UniGene clone set RZPD 1 Build 138, NCBI ), which consisted of ~33,792 cDNA clone inserts spotted in duplicate onto membranes . These microarrays (n = 31) were hybridized with 33P-labeled cDNA derived from total RNA extracted from biopsy material from the sigmoidal colon of normal (control, n = 11), and patients with Crohn's disease (condition A, n = 10) and ulcerative colitis (condition B, n = 10). To emphasize that our normalization technique can be used to normalize other array systems, the second array set used was a smaller, but widely used, commercially available microarray system. Data set B (Table 1) was generated by using Atlas Rat cDNA microarrays (Clontech, 588 genes) probed with rat brain tissue, from control (cerebellum n = 10, olive n = 10) and harmaline treated (cerebellum n = 10, olive n = 9) animals. A third microarray data set, data set H (Table 1) was included to demonstrate the normalization method on two channel fluorescent based (Cy3/Cy5) oligonucleotide systems. These custom produced boutique microarrays (n = 5) contained 1024 spots, and were used in a study to identify differences in apoptotic mechanisms in two different mouse cell lines. Microarrays were probed according to established protocols and exposed to imaging plates overnight (BAS-MS 2325) and scanned at a 50 μm resolution on a FLA-3000G phosphoimager (Raytest, Germany). Image gridding was carried out using VisualGrid® software , and intensity data was stored in a relational database and normalized and analyzed using database stored procedures and Perl scripts. All data was normalized from raw data, no background subtraction or other inter-array normalization was performed. Plots were generated using the Grace software package .
Normalization was accomplished by transforming the data such that the coefficient c and proportionality of the Zipf's power function (formula 1) are identical for all microarrays. This is easily achieved using a regression model on the loge intensity versus loge rank transformed data, which has the general form,
ln (y) = a + b ln (r) + e (2)
where y is the intensity, r is the rank, a is the regression constant (corresponding to proportionality in Zipf's power function), b is the regression coefficient (corresponding to the coefficient c in Zipf's power function), and e is an error coefficient, which is assumed to be normally distributed.
The first step in this three step procedure was to compute the median intensity of each gene over all microarrays to establish ranks, which were used as the 'reference' to which all microarrays were normalized. This was done by taking the median intensity (y med ) of each gene, over all microarrays on which it was measured, and sorting the resulting list of medians to obtain their median ranks (r med ). The regression model (2) is applied to the loge median intensities and their ranks to estimate a med and b med using the least squares method,
The ranking of genes by their median intensities effectively groups genes of similar overall expression level along the log rank axis. Under the assumptions that most genes are not differentially expressed, the reference curve generated from the median intensities should have an identical regression coefficient and constant to that of each individual microarray plotted using the ranks determined by the medians. For the genes which are differentially expressed, the median value represents a 'center' around which expression levels on each individual array may vary, and the neighbouring (by rank) genes, which do not (or only slightly) vary, act to stabilize the regression line and allow normalization to be performed.
In the second step of the normalization procedure, the regression model was applied individually to each microarray using the same ranking as the reference curve,
This results in a set of coefficients a k and b k which are estimated individually for each array using the least squares method, where k is equal to the number of microarrays in one channel systems, and equal to 2 time the number of microarrays (one for each channel) in two channel systems. Data from two channel arrays were treated in the same way as one channel systems, i.e. each channel was treated independently.
In the third step, the difference between the expected gene intensity value on the k th array and that of the reference curve was applied as the normalization factor,
A scaling factor was applied to the raw data before normalization such that the values y k , and were always greater than one to avoid negative values after log transformation. After normalization, the same scaling factor was applied to the data to back transform to their original magnitude. For example, if the smallest raw value in the data set was 0.1, the unlogged raw data was multiplied by a scaling factor of 10 before normalization, and the unlogged normalized data was divided by the same scaling after normalization.
In the special case of our third microarray data set (see Methods: Data Acquisition) which was a boutique array, the same procedure as described above was applied with the following modifications. Each microarray contained 32 spots each of internal positive controls (GAPDH, glyceraldehyde-3-phosphate_dehydrogenase) and internal negative controls (spotting buffer). The medians of all gene intensities were computed (including internal positive and negative controls), and median ranks were assigned as described. However, only the medians of the 64 internal control spots were used to estimate amed and bmed, and only the 64 internal control spots from each array were used to estimate ak and bk. In both cases, the ranks generated from the entire data set, were used. The normalization factor was then applied over the entire data set as described above.
An alternative to the used of internal control spots for the normalization of boutique microarrays was also explored. Wilson, et. al.  described a method wherein a set of 'housekeeping' genes is selected a priori from the data set by virtue of their low variance in intensity and such that the entire range of intensities observed on the microarrays is uniformly represented. We also applied the Zipf's law normalization technique to our boutique microarrays using the set of housekeeping genes selected using the method of Wilson, et. al.
In addition to the normalization method based on Zipf's law, all data sets were normalized to a global mean (the mean of logged intensities from all microarrays) and the quantile method. The quantile method is applied by ranking the genes in each array by intensity, taking the median intensity at each rank, and replacing each gene intensity with the median intensity corresponding to the same rank. All normalization methods were compared to each other and to the raw data distribution using box plots and MA plots (pairwise array comparisons of the log-intensity ratio (M) to the mean log-intensity (A)). The two channel boutique microarray data set allowed further normalization methods not possible on one channel array systems to be applied. We normalized this data set using the popular loess method , and a modified Loess method specifically designed for boutique arrays using selected housekeeping genes described by Wilson, et. al. .
The Zipf's normalization procedure was initially implemented as an SQL stored procedure in a relational database. However, because this is not easily transferable to other systems, we provide two further implementations, a Perl script and an Excel macro [see Additional files 4, 5]. Implementations are available for download from our website  and as additional files accompanying this paper. Both the Perl script and Excel macro implement matrix algebra style computation, using either built-in functions or the Perl PDL module . Normalization of two channel arrays with the loess method was performed using the marray package from R's Bioconductor . Loess normalization using selected housekeeping genes and the selection of the housekeeping genes themselves was done with the tRMA package  which is publicly available for download on the internet. Sample data sets are also provided with this paper [see Additional files 6, 7, 8].
Normalization method comparison
To compare and evaluate the effectiveness of the various normalization methods applied in this paper, several well established methods were used along with some less common techniques. MA plots  are a convenient way to examine differences in fluorescent marker efficiency and other dye effects in two channel microarray systems. In addition to the standard practice of generating within-array MA plots, we apply them additionally to one channel systems and between arrays in two channel systems to evaluate the extent to which a normalization procedure allows for multiple pairwise comparisons between microarrays. Plots of loge intensity versus loge rank fitted with linear regressions are a way to visually evaluate the normalization procedure according to the criteria of the Zipf's Law normalization. Specifically, all arrays have identical coefficients c and proportionality for the Zipf's power function when the slops and y-intercepts of the regression lines are identical. Finally, to quantify the similarity between microarray distributions after normalization, the mean Kolmogorov-Smirnov value was calculated over all possible pairwise combinations of microarrays within an experiment. In the case of two channel arrays, the mean of within-array Kolmogorov-Smirnov values was also computed (n = the number of arrays). It should be emphasized that even though the Kolmogorov-Smirnov values are technically a test statistic, no statistical test is performed. The values are here used only as a measure of similarity between microarray distributions.
Microarray platform comparison
The underlying premise of the Zipf's normalization method is that microarray data distributions follow a power law distribution such that the relationship between the log intensities and the log ranks is clearly linear. While this assumption holds true for the three data sets we present in this paper, to evaluate the general applicability of the method we also examined eight publicly available data sets (Table 1, data sets C-G, I, K-L) from the NCBI Gene Expression Omnibus , and one unpublished data set from an independently maintained website  (Table 1, data set J). The survey contains a variety of microarray system types (cDNA vs. Oligo based, radioactivity vs. dye labeled systems, academic vs. commercially produced) and two SAGE experiments for comparison. Two plots were generated for each data set to ascertain the conformity to the Zipf's power law distribution and the log normal distribution respectively. For each data set, a representative array was constructed by ranking the intensities within each array, and then mean over ranks were taken. To determine how well data sets follow the Zipf's power law distribution, log intensity vs. log rank plots were constructed and linear regressions were performed. Data distributions, which were very linear in form, closely follow the power law distribution. A second plot of the distribution of (log y – μ) / σ, where y is the mean intensity over ranks, and μ and σ2 are the mean and variance, was made for each data set to visualize the conformity to log normal distribution.
List of abbreviations
Expressed Sequence Tag
log ratio (M) vs. mean log intensity (A)
National Center for Biotechnology Information
Deutsches Ressourcenzentrum für Genomforschung GmbH
Serial analysis of gene expression
Structured Query Language
The authors wish to thank Alexander Zein and Carl Manaster for critical input on data analysis techniques. We would like to express our appreciation to the clinicians and volunteers who provided tissue samples, and Hans Moises and Henrik Wilms for rat brain samples. We gratefully acknowledge the technical assistance of Brigitte Mauracher, and the invaluable assistance of the Max-Planck Institute for Molecular Genetics in Berlin, in particular Hans Lehrach, Holger Eickhoff and Elke Rohlfs. We also thank Sandra Freitag for advice on formulating the equations. This research was supported in part by a Training and Mobility of Researchers (TMR) grant, as well as grants from the German National Genome Research Program, the National Genome Research Network (NGFN) and the DFG (FOR423).
- Brown PO, Botstein D: Exploring the new world of the genome with DNA microarrays. Nat Genet 1999, 21(1 Suppl):33–37. 10.1038/4462View ArticlePubMedGoogle Scholar
- Lander ES: Array of hope. Nat Genet 1999, 21(1 Suppl):3–4. 10.1038/4427View ArticlePubMedGoogle Scholar
- Tsodikov A, Szabo A, Jones D: Adjustments and measures of differential expression for microarray data. Bioinformatics 2002, 18(2):251–260. 10.1093/bioinformatics/18.2.251View ArticlePubMedGoogle Scholar
- Wilson DL, Buckley MJ, Helliwell CA, Wilson IW: New normalization methods for cDNA microarray data. Bioinformatics 2003, 19(11):1325–1332. 10.1093/bioinformatics/btg146View ArticlePubMedGoogle Scholar
- Yang YH, Dudoit S, Luu P, Lin DM, Peng V, Ngai J, Speed TP: Normalization for cDNA microarray data: a robust composite method addressing single and multiple slide systematic variation. Nucleic Acids Res 2002, 30(4):e15. 10.1093/nar/30.4.e15PubMed CentralView ArticlePubMedGoogle Scholar
- Zien A, Aigner T, Zimmer R, Lengauer T: Centralization: a new method for the normalization of gene expression data. Bioinformatics 2001, 17(Suppl 1):S323–31.View ArticlePubMedGoogle Scholar
- Velculescu VE, Madden SL, Zhang L, Lash AE, Yu J, Rago C, Lal A, Wang CJ, Beaudry GA, Ciriello KM, Cook BP, Dufault MR, Ferguson AT, Gao Y, He TC, Hermeking H, Hiraldo SK, Hwang PM, Lopez MA, Luderer HF, Mathews B, Petroziello JM, Polyak K, Zawel L, Kinzler KW, et al.: Analysis of human transcriptomes. Nat Genet 1999, 23(4):387–388. 10.1038/70487View ArticlePubMedGoogle Scholar
- Raychaudhuri S, Stuart JM, Altman RB: Principal components analysis to summarize microarray experiments: application to sporulation time series. Pac Symp Biocomput 2000, 455–466.Google Scholar
- Kerr MK, Martin M, Churchill GA: Analysis of variance for gene expression microarray data. J Comput Biol 2000, 7(6):819–837. 10.1089/10665270050514954View ArticlePubMedGoogle Scholar
- Friedman N, Linial M, Nachman I, Pe'er D: Using Bayesian networks to analyze expression data. J Comput Biol 2000, 7(3–4):601–620. 10.1089/106652700750050961View ArticlePubMedGoogle Scholar
- Furusawa C, Kaneko K: Zipf's law in gene expression. Phys Rev Lett 2003, 90(8):088102. Epub 2003 Feb 26. 10.1103/PhysRevLett.90.088102View ArticleGoogle Scholar
- Ogasawara O, Kawamoto S, Okubo K: Zipf's law and human transcriptomes: an explanation with an evolutionary model. C R Biol 2003, 326(10–11):1097–1101.View ArticlePubMedGoogle Scholar
- Ueda HR, Hayashi S, Matsuyama S, Yomo T, Hashimoto S, Kay SA, Hogenesch JB, Iino M: Universality and flexibility in gene expression from bacteria to human. Proc Natl Acad Sci U S A 2004, 101(11):3765–9. Epub 2004 Mar 03. 10.1073/pnas.0306244101PubMed CentralView ArticlePubMedGoogle Scholar
- Hoyle DC, Rattray M, Jupp R, Brass A: Making sense of microarray data distributions. Bioinformatics 2002, 18(4):576–584. 10.1093/bioinformatics/18.4.576View ArticlePubMedGoogle Scholar
- Zipf GK: The psycho-biology of language; an introduction to dynamic philology. Boston, , Houghton Mifflin Company; 1935:ix , 2 *., -336.Google Scholar
- Schuchhardt J, Beule D, Malik A, Wolski E, Eickhoff H, Lehrach H, Herzel H: Normalization strategies for cDNA microarrays. Nucleic Acids Res 2000, 28(10):E47. 10.1093/nar/28.10.e47PubMed CentralView ArticlePubMedGoogle Scholar
- Bolstad BM, Irizarry RA, Astrand M, Speed TP: A comparison of normalization methods for high density oligonucleotide array data based on variance and bias. Bioinformatics 2003, 19(2):185–193. 10.1093/bioinformatics/19.2.185View ArticlePubMedGoogle Scholar
- Edgar R, Domrachev M, Lash AE: Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res 2002, 30(1):207–210. 10.1093/nar/30.1.207PubMed CentralView ArticlePubMedGoogle Scholar
- Dudoit S, Yang YH, Callow MJ, Speed TP: Statistical methods for identifying differentially expressed genes in replicated cDNA microarray experiments. Statistica Sinica 2002, 12(1):111–139.Google Scholar
- Kendall MG, Stuart A, Ord JK: Tests of fit based on the sample distribution function: Kolmogorov’s Dn. In Kendall's advanced theory of statistics. Volume 2. Fifth edition. New York , Oxford University Press; 1987:1187–1188.Google Scholar
- Wentian Li’s literature review of Zipf’s Law[http://www.nslij-genetics.org/wli/zipf/index.html]
- Benford F: The Law of Anomalous Numbers. Proc Am Philos Soc 1936, 78: 551–572.Google Scholar
- Kroll TC, Wolfl S: Ranking: a closer look on globalisation methods for normalisation of gene expression arrays. Nucleic Acids Res 2002, 30(11):e50. 10.1093/nar/30.11.e50PubMed CentralView ArticlePubMedGoogle Scholar
- Kuznetsov VA, Knott GD, Bonner RF: General statistics of stochastic process of gene expression in eukaryotic cells. Genetics 2002, 161(3):1321–1332.PubMed CentralPubMedGoogle Scholar
- Schadt EE, Li C, Ellis B, Wong WH: Feature extraction and normalization algorithms for high-density oligonucleotide gene expression array data. J Cell Biochem Suppl 2001, Suppl(37):120–125. 10.1002/jcb.10073View ArticleGoogle Scholar
- Website of the Deutsches Ressourcenzentrum für Genomforschung[http://www.rzpd.de]
- Homepage of GPC Biotech, makers of VisualGrid®[http://www.gpc-biotech.com]
- Grace plotting software[http://plasma-gate.weizmann.ac.il/Grace]
- Original data sets and Zipf’s normalization software[http://www.mucosa.de/zipfs/zipfs_normalization.html]
- The Perl Data Language homepage[http://pdl.perl.org]
- Ross DT, Scherf U, Eisen MB, Perou CM, Rees C, Spellman P, Iyer V, Jeffrey SS, Van de Rijn M, Waltham M, Pergamenschikov A, Lee JC, Lashkari D, Shalon D, Myers TG, Weinstein JN, Botstein D, Brown PO: Systematic variation in gene expression patterns in human cancer cell lines. Nat Genet 2000, 24(3):227–235. 10.1038/73432View ArticlePubMedGoogle Scholar
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.