Resolution of large and small differences in gene expression using models for the Bayesian analysis of gene expression levels and spotted DNA microarrays
© Townsend; licensee BioMed Central Ltd. 2004
Received: 13 February 2004
Accepted: 05 May 2004
Published: 05 May 2004
The detection of small yet statistically significant differences in gene expression in spotted DNA microarray studies is an ongoing challenge. Meeting this challenge requires careful examination of the performance of a range of statistical models, as well as empirical examination of the effect of replication on the power to resolve these differences.
New models are derived and software is developed for the analysis of microarray ratio data. These models incorporate multiplicative small error terms, and error standard deviations that are proportional to expression level. The fastest and most powerful method incorporates additive small error terms and error standard deviations proportional to expression level. Data from four studies are profiled for the degree to which they reveal statistically significant differences in gene expression. The gene expression level at which there is an empirical 50% probability of a significant call is presented as a summary statistic for the power to detect small differences in gene expression.
Understanding the resolution of difference in gene expression that is detectable as significant is a vital component of experimental design and evaluation. These small differences in gene expression level are readily detected with a Bayesian analysis of gene expression level that has additive error terms and constrains samples to have a common error coefficient of variation. The power to detect small differences in a study may then be determined by logistic regression.
Spotted DNA microarrays can be used to measure genome-wide gene expression levels in cells of different genotypes, in different developmental states, or within different environments. The precision and accuracy of these measurements depend on the technical performance of the microarray, the degree of replication of the experiment, and the suitability of the model used to analyze the data. A number of models have been advanced for the statistical analysis of experimental designs involving two samples [1–4]. Two methods, a classical ANOVA method [5–7] and a Bayesian method , have been designed for the analysis of experimental designs involving multiple nodes of expression such as genotypes, environments, and developmental states. These analyses yield quantitative results on the expression level of a gene, evaluating data from direct hybridizations as well as data from hybridizations that are informative through transitive inference .
Optimal statistical inference depends upon the choice of model used for analysis. Townsend and Hartl  derived a core model that has been widely used for the estimation of gene expression levels and statistical significance in multifactorial experiments (e.g. [9–15]). This model assumed additive small error terms and either estimated error variances for each genotype, environment, or developmental state or estimated a single variance for all genotypes, environments, or developmental states. The ANOVA models of Kerr et al.  and Wolfinger et al.  have also been widely used, and assume multiplicative small error terms. Accordingly, Bayesian models are considered here that incorporate multiplicative error terms. A number of other studies have correlated error variance to raw expression level [2, 16–18]. To evaluate the potential effect of this correlation on ratio measurements, models are developed here that constrain the relationship between the error variances and the expression levels to a constant coefficient of variation. Nested error models for spotted DNA microarrays are compared using the Bayesian information criterion for model choice .
The power to detect differences in gene expression using these models is evaluated, and the relationship between the estimated expression level, the number of replicate hybridizations, and the ability to determine the statistical significance of small differences in gene expression is explored both for simulated and empirical data. A summary statistic for determining the fold-resolution detectable as significant in empirical microarray studies is presented.
Model with small additive error effects
The intensity of hybridization of a DNA spot on a microarray is often used as measure of gene expression, but the raw intensity is subject to a number of confounding error terms, such as DNA concentration in a spot and sequence hybridization efficiency. As foreseen by the pioneers of DNA microarray technology, these confounding effects (regardless of their multiplicative or additive nature) are eliminated by consideration of the ratio of hybridization of two samples . The remaining small error terms may be modeled either additively or multiplicatively.
Townsend and Hartl  modeled them additively, deriving a density function for the observed ratio of gene expression in the i th and j th condition, z ij , as
where μ i is the expression level in condition i and σ2 i is the variance in condition i. If n conditions or genotypes are under study, direct use of this likelihood function requires estimation of 2n - 1 parameters (n - 1 expression levels plus n variances) for each gene.
One alternative is to constrain the variance so that it is common and equal among nodes of an experimental design , thus reducing the number of parameters requiring estimation to n (i.e. n - 1 expression levels plus one variance). Another alternative is to constrain the variances such that they have a consistent relationship to the expression levels in each node. For example, conditions or genotypes can have a common coefficient of variation (CV) for each gene, ν = σ i / μ i for all i. This alternative also requires the estimation of n parameters (n - 1 expression levels and one CV). It is intuitively motivated by the consideration that larger means for a population are accompanied by larger variances across many phenomena in the sciences . In order to implement a model in which all nodes of an experimental design have a common error CV for each gene, equation 1 may be rewritten, substituting ν2μ i 2 for σ i 2:
Equation 2 can then be used with a prior to construct a Markov chain whose stationary distribution is the posterior distribution of the parameters given the data [22, 23], in all other ways following the algorithm of Townsend and Hartl . The formulation in Equation 2 has additional appeal over Equation 1 when used (as it will be below) within a Markov Chain Monte Carlo (MCMC) analysis. Because values of μ and ν tend to scale similarly across the real line compared to μ and σ2, less tuning of the MCMC jump size may be necessary to achieve a satisfactorily mixed chain.
Model with small multiplicative error effects
An alternative to additively modeled error is to model error multiplicatively, such that the post-normalization intensity in one fluorescence channel at a reporter spot is
where μ is the absolute quantity of mRNA per cell, the c m are spot-associated terms of arbitrary distribution for any a multiplicatively confounding factors, the c l are spot-associated terms of arbitrary distribution for any q-t linearly confounding factors, and ε is a term for small random errors not associated with the spot. The observed ratios of intensities after normalization, y ij , would then be
Taking the log of both sides,
The formulation in Equation 4 has some evident similarities to formulations in ANOVA models of gene expression measurement error, where the confounding terms c correspond to the array spot effects identified by Kerr et al.  and Wolfinger et al. , except that the derivation presented here does not assume that these terms are lognormally distributed. Note that these confounding terms are generally not of biological interest and can immediately cancel in equations 3 or 4. Assuming that error terms log ε i and log ε j are composed of many small, unbiased effects, and scaling them so that they are distributed with variances specific to each node, σ i 2 and σ j 2, it follows from equation 3 that the ratio data, z ij , are drawn from a ratio of two lognormal distributions. The numerator is drawn from a lognormal distribution with parameters μ i and σ i 2, and the denominator is drawn from a lognormal distribution with parameters μ j and σ j 2. Just as the difference of two Gaussians is itself Gaussian, the ratio of two lognormals is lognormal, thus the probability density function is
Following in all other ways the algorithm of Townsend and Hartl , Equation 5 can then be used with a prior to construct a Markov chain, the stationary distribution of which is the posterior distribution of the parameters given the data. Furthermore, just as with equation 2, all variances for each node may be constrained to be equal or each may be constrained to be linearly proportional to its respective expression level by a single CV. In the latter case, with ν = σ i / μ i = σ j / μ j for all i and all j,
The three-dimensional matrix of ratio results from DNA microarray comparisons, Z, may be constructed, with dimensions i denoting the sample labeled with one fluorophore, j denoting the sample labeled with another, and k denoting the replicate ordinate of that particular dye-labeled comparison. Then, for any continuous structure of comparisons among the nodes of interest, the likelihood density for the parameters μ l and ν l , 1 ≤ l ≤ n, is, by Bayes' rule,
where g(μ i , ν i , μ j , ν i ) is the prior distribution of the parameters, and where the probability f(z ijk ) of empty elements in the data matrix Z is properly evaluated as one.
Appropriate informative priors for the variance of microarray data are under investigation [2, 4, 24]. In this paper, a noninformative prior distribution, uniform across positive real numbers, has been used for both the expression levels and for their variances and CVs. The range has been nominally constrained between zero 100, though that upper constraint makes no difference for the datasets examined here. The uniform prior gives the microarray data itself the greatest impact on the inferred expression levels and variances, and implies that credible intervals around parameter estimates (the Bayesian equivalents of classical confidence intervals) are close to those that would be found by maximum likelihood.
Fortunately, we may use the constant denominator of the Bayes' rule formulation (Equation 5) to assert that
Our proposed values are constructed in two separate steps. First, two of the n gene expression level parameters from are chosen at random. A step size is drawn at random from a triangular distribution centered at zero with range [-Δμ, +Δμ]. The first of the two chosen parameters is incremented by the chosen step size, and the second is decremented by the same quantity, so that is maintained, where the apostrophe indicates a proposed parameter value. In the next iteration, each of the CV parameters in is separately incremented by an amount drawn at random from a triangular distribution with range [-Δν, +Δν] to form . The conjecture is accepted for the next state of the Markov chain if
Otherwise the original state is retained for the next iteration of the Markov Chain.
This implementation of these models can accommodate complex experimental designs, where a number of genotypes, environments, and developmental time points are examined. Within this framework, missing data (e.g. excluded single spots, or even missing hybridizations) do not require special consideration or a change in methodology; credible intervals and P values reflect accurately the degree to which the data informs each estimate. This software allows the quantitative information on gene expression levels from microarrays to be thoroughly analyzed and carefully considered in assessing the biological effects of genetic or environmental differences of cellular state.
Output from the software implementation is in the form of a tab-delimited text file with one header row. Each row thereafter displays the results for a single gene, including columns with: the estimate of expression level for each node (the median of the posterior distribution); the additions and subtractions to make 95% upper and lower bounds on that estimate; the stationary acceptance rates for the Monte Carlo steps for that gene; and the posterior probabilities (P values) for whether the expression level of a gene in each expression node is greater, or lesser, than the expression level of that gene in each other expression node.
Nested model choice
The common variance (AV, MV) and common CV (AC, MC) models are both nested within their respective general unconstrained variance model (AU, MU). The same number of parameters is estimated in both of the nested models. They differ only in how the estimated variances are constrained with relation to the estimated expression levels. Whether the nested models are appropriate compared to the general model may be assessed using the Bayesian Information Criterion , which is to choose the model m that maximizes
where M m is the maximum likelihood of model m, h m is the number of parameters estimated in the model, and n is the number of observations.
Tests of power
For each distribution, 1000 measurements of gene expression level were simulated where both samples had the same expression level, and one hundred measurements were simulated for ratios of expression level of 1.1, 1.25, 1.5, and 2. Variance and CV parameters for all the above distributions simulated expression levels were set at the average values inferred from the dataset of Townsend et al.  under additive models. Note that, although parameters of each distribution were generally chosen so that the variances of the ratio output of each distribution would be similar, no attempt was made to make higher moments than the mean identical. Therefore, the relevant comparisons are between analysis methods on a given simulated dataset, and frequencies of significance calling are not directly comparable across simulated datasets.
Power to detect a difference in gene expression depends critically on the true factor of fold-difference between samples. A continuous logistic function,
describing the probability of detection of statistical significance, p, of simulated log2 factors of difference in gene expression, x, was parameterized with an intercept, b, and slope, m, by logistic regression. The same regression was performed on real data by substituting estimates of the factor of difference in gene expression level for known factors of difference, thus providing a profile of the power of an experiment to detect differences in gene expression. A useful metric for such an analysis is the factor of difference in gene expression level that has a fifty percent chance of being identified as significant. Herein, this is referred to as the GEL50, for the Gene Expression Level at which there is a 50 percent chance of detection of statistical significance.
Analysis of general, unconstrained variance vs. nested, constrained variance models
s / gene**
Townsend et al. (2003)
Townsend et al. (2003)
Townsend et al. (2003)
Townsend et al. (2003)
Townsend et al. (2003)
Townsend et al. (2003)
Sudarsanam et al. (2000)
Sudarsanam et al. (2000)
Sudarsanam et al. (2000)
Sudarsanam et al. (2000)
Sudarsanam et al. (2000)
Sudarsanam et al. (2000)
Computation time for analysis of published data sets varied across models (Table 1). Computation using additive models (AV, AC, AU) was more rapid than computation using multiplicative models. Regardless of whether small error terms were modeled as additive or multiplicative, constrained CV models (AC, MC) were faster than constrained variance (AV, MV) or general unconstrained (AU, MU) models. Furthermore, the relative ranks of these models in terms of speeds of computation, without exception, remained as above in all analyses of simulated datasets.
Higher power to detect true differences, although important in practice for the purpose of choice of model in an experimental study, does not indicate a better fit to the data. This is made clear by comparing Figure 1A with Figure 1C, and by comparing Figure 2B with Figure 2D. If the increased power to detect true differences in gene expression were solely due to a better fit to the data, then analysis via model AV would outperform analysis via model AC in Figure 1A, and analysis via model MV would outperform analysis via model MC in Figure 2B. Thus, constraining the nodes of the experimental design to a single error CV yields greater power than constraining the nodes to single error variance, regardless of which analysis model fits the data better.
In comparison to performing logistic regression of the frequency of positive calls versus true differences in gene expression level, a logistic regression of the frequency of positive calls versus the estimates of gene expression level derived from analysis of the simulated data may be performed (Figure 4B). Stochasticity in the estimation due to small sample size causes dispersion of the values along the abscissa. Nevertheless, this regression on estimates of gene expression level shows the same decreasing probability of significant call with decreasing gene expression difference between nodes of an experimental design.
Distinguishing the optimal models to use for the analysis of replicated spotted DNA microarray data is important. Optimized models will yield qualitatively more accurate lists of significantly differently expressed genes, and quantitatively more precise resolution of smaller differences in gene expression. The Bayesian Information Criterion for model selection can be used to choose between models that invoke distinct error variances or coefficients of variation for each node as characterized by genotype, environment, and developmental state, and the nested models that invoke a single variance or CV for all nodes. The values of the BIC for the relatively small studies examined here (Table 1) clearly support analysis with the nested models that invoke a single variance or CV.
In addition to direct assessment of the fit of the model to the data, power to detect known differences may guide model choice. Generally, the ranking of the power of models was consistent regardless of the distribution used to simulate the data (Figures 1, 2, and 3). For unconstrained models that estimate a variance or CV term for each node, the analysis model incorporating an assumption of small multiplicative error terms (model MU) had greater power to detect differences than the analysis model incorporating an assumption of small additive error terms (model AU). Nested models that invoke a single variance or CV had higher power to detect known differences when the analysis model incorporated an assumption of small additive error terms (Equation 1, i.e. models AC and AV had higher power than model AU.) Among analysis models incorporating an assumption of small multiplicative error terms (Equation 2), only one of the two nested models, model MC, had consistently higher power than the unconstrained model. Overall, a model incorporated an assumption of small additive error terms and a single error CV for all nodes (model AC) had the greatest power to detect differences in gene expression level. In practice, model AC was also the fastest computationally (Table 1), perceptibly requiring fewer tuning steps to find an appropriate jump size for the generation of posterior distributions by Markov Chain Monte Carlo.
If variances are generally proportional to their expression levels, then the constrained CV models (AC and MC) pertain. A linear regression of the estimated coefficients of variation to their respective expression levels should have positive slope. Specifically, regressions on the datasets here typically have positive slope (y = ~ 0.4x + c) and are highly statistically significant, although the data exhibit considerable scatter and thus poor correlation (r2 ~ 0.04). These data sets are barely large enough to estimate error variances in a gene-by-gene manner using the general model. Future experimental data with greater replication, analyzed by the general model, will yield higher precision estimates of the error variances and thus better resolution of this question.
When the nominal false positive rate α = 0.05, all models have an actual false positive rate that is moderately to considerably less than 0.05, averaging 0.02 (Figures 1, 2, and 3). These false positive rates are close enough to zero that precise estimation of the frequency requires extensive computation, but generally, the false positive rate was slightly higher with the more powerful models. The slightly lower than nominal false-positive rate is due to the flat prior on the error variance or error CV, that is slightly too permissive of large estimates of the variance and of the CV. Further investigation into the use of a more informative prior distribution (such as the gamma distribution, e.g. ), is called for, requiring larger studies with greater replication.
Prediction of the number of replicates required for statistical significance testing of microarray data is theoretically possible [29, 30], by making specific assumptions about the error variances and the level of gene expression difference of interest. Here, empirical examination of the power to detect significant differences at different gene expression levels in different studies (Figure 5) has the potential to simply and rapidly convey vital evaluative feedback about the design, replication, and technical performance of a set of hybridizations. All three of these factors contribute to the ability of a microarray study to resolve small factors of difference in gene expression between nodes in an experimental design. This ability to resolve differences can be summarized by a single, intuitive parameter: the gene expression level at which there is a 50% frequency of significant calls (GEL50). Note, however, that the regressions in Figure 5, like the volcano plots of Wolfinger et al. , show genes with a broad range of estimated expression levels that are significant, and genes with a broad range of estimated expression levels that are not significant. Therefore, analyses that invoke a fold-change threshold as an indication of significance should be avoided.
From the datasets analyzed here, it is clear that increased replication leads to greater resolution of small differences in gene expression (Figure 5). This small number of studies, of varying technical quality, does not warrant a strictly quantitative empirical formula for the GEL50 based on the number of nodes in the experimental design and the number of replicate hybridizations. However, a very crude rule of thumb based upon examination of the quality and resolution of these and other datasets is that the GEL50 resolution of a study is of the form , where n is the number of nodes in the design and r is the total number of hybridizations performed. The studies examined here all contained replicated comparisons, and, in accord with MIAME standards , reported ratio results from each hybridization. Future analyses of a range of additional studies that also report results of each hybridization for each gene will have the potential to reveal a more accurate and precise prediction of power using more sources of information about the quality of the microarray hybridizations and about the optimal design of multifactorial experiments .
Increased power to detect differences in gene expression, consequent to better analysis, better replication, or better technical performance, identifies more significant differences in gene expression of genes with smaller and smaller true expression differences. These small differences in gene expression are not only present [10, 32], they are relevant to the evolution of gene regulation  and to organismal function and phenotype [32, 33]. Transcription factors, for instance, may have enormous impact on cellular function with minimal changes in expression level [34, 35]. The detection of the differential expression of transcription factors is often a major goal of many microarray studies. Therefore, understanding the resolution of difference in gene expression that is detectable as significant is a vital component of experimental design and evaluation.
Availability and requirements
Project name: Bayesian Analysis of Gene Expression Level (BAGEL)
Project home page: http://plantbio.berkeley.edu/~taylor/jto.html
Operating system(s): MacOS 9, MacOS X, Windows and Linux.
Other requirements: none
Model abbreviations and relations
- Models used will be abbreviated with a two-letter acronym. The first letter indicates (A)dditive or (M)ultiplicative error:
and the second letter indicates a general (U)nconstrained variance model, a constrained (V)ariance model, or a constrained (C)oefficient of variation model. Thus, the AV and AC models are nested within the AU model, while the MV and MC models are nested within the MU model. With n nodes in the experimental design, the AU and MU models both have 2n - 1 parameters (n - 1 expression levels plus n variances), and the AV, AC, MV, and MC models all have n parameters (n- 1 expression levels, plus 1 variance or CV).
Thanks to Takao Kasuga, Alison Galvani, and Betty Gilbert for comments on the manuscript. JPT was supported while performing this work by a Research Fellowship from the Miller Institute for Basic Research in Science.
- Ideker T, Thorsson V, Siegel AF, Hood LE: Testing for differentially-expressed genes by maximum-likelihood analysis of microarray data. Journal of Computational Biology 2000, 7: 805–817. 10.1089/10665270050514945View ArticlePubMedGoogle Scholar
- Baldi Pierre, Long Anthony: A Bayesian framework for the analysis of microarray expression data: regularized t-test and statistical inferences of gene changes. Bioinformatics 2001, 17: 509–519. 10.1093/bioinformatics/17.6.509View ArticlePubMedGoogle Scholar
- Theilhaber Joachim, Bushnell Steven, Jackson Amanda, Fuchs Rainer: Bayesian estimation of fold-changes in the analysis of gene expression: the PFOLD algorithm. Journal of Computational Biology 2001, 8: 585–614. 10.1089/106652701753307502View ArticlePubMedGoogle Scholar
- Tseng George C., Oh Min-Kyu, Rohlin Lars, Liao James C., Wong Wing Hung: Issues in cDNA microarray analysis: quality filtering, channel normalization, models of variations and assessment of gene effects. Nucleic Acids Research 2001, 29: 2549–2557. 10.1093/nar/29.12.2549PubMed CentralView ArticlePubMedGoogle Scholar
- Kerr M. Kathleen, Martin Mitchell, Churchill Gary A.: Analysis of variance for gene expression microarray data. Journal of Computational Biology 2000, 7: 819–837. 10.1089/10665270050514954View ArticlePubMedGoogle Scholar
- Kerr M. Kathleen, Churchill Gary A.: Experimental design for gene expression microarrays. Biostatistics 2001, 2: 183–201. 10.1093/biostatistics/2.2.183View ArticlePubMedGoogle Scholar
- Wolfinger Russell D., Gibson Greg, Wolfinger Elizabeth D., Bennett Lee, Hamadeh Hisham, Bushel Pierre, Afshari Cynthia, Paules Richard S.: Assessing gene significance from cDNA microarray expression data via mixed models. Journal of Computational Biology 2001, 8: 625–637. 10.1089/106652701753307520View ArticlePubMedGoogle Scholar
- Townsend Jeffrey P., Hartl Daniel L.: Bayesian analysis of gene expression levels: statistical quantification of relative mRNA level across multiple treatments or samples. Genome Biology 2002, 3: research0071.1–71.16. 10.1186/gb-2002-3-12-research0071View ArticleGoogle Scholar
- Townsend Jeffrey P.: Multifactorial experimental design and the transitivity of ratios with spotted DNA microarrays. BMC Genomics 2003, 4: 41. 10.1186/1471-2164-4-41PubMed CentralView ArticlePubMedGoogle Scholar
- Townsend Jeffrey P., Cavalieri Duccio, Hartl Daniel L.: Population genetic variation in genome-wide gene expression. Molecular Biology and Evolution 2003, 20: 955–963. 10.1093/molbev/msg106View ArticlePubMedGoogle Scholar
- Meiklejohn Colin D., Parsch John, Ranz JM, Hartl Daniel L.: Rapid evolution of male-biased gene expression in Drosophila. Proceedings of the National Academy of Sciences - USA 2003, 100: 9894–9899. 10.1073/pnas.1630690100View ArticleGoogle Scholar
- Ranz JM, Castillo-Davis CI, Meiklejohn Colin D., Hartl Daniel L.: Sex-dependent gene expression and evolution of the Drosophila transcriptome. Science 2003, 300: 1742–1745. 10.1126/science.1085881View ArticlePubMedGoogle Scholar
- Silverman Neal, Zhou Rui, Ehrlich Rachel, Hunter Mike, Bernstein Erik, Schneider David, Maniatis Tom: Immune activation of NF-kB and JNK requires Drosophila TAK1. The Journal of Biological Chemistry 2003, 278: 48928–48934. 10.1074/jbc.M304802200View ArticlePubMedGoogle Scholar
- Whitfield Charles W., Cziko Anne-Marie, Robinson Gene E.: Gene expression profiles in the brain predict behavior in individual honey bees. Science 2003, 302: 296–299. 10.1126/science.1086807View ArticlePubMedGoogle Scholar
- Grozinger Christina M., Sharabash Noura M., Whitfield Charles W., Robinson Gene E.: Pheromone-mediated gene expression in the honey bee brain. Proceedings of the National Academy of Sciences - USA 2003, 100: 14519–14525. 10.1073/pnas.2335884100View ArticleGoogle Scholar
- Rocke David M., Durbin Blythe: A model for measurement error for gene expression arrays. Journal of Computational Biology 2001, 8: 557–569. 10.1089/106652701753307485View ArticlePubMedGoogle Scholar
- Mutch DM, Berger A, Mansourian R, Rytz A, Roberts MA: The limit fold change model: a practical approach for selecting differentially expressed genes from microarray data. BMC Bioinformatics 2002, 3: 17. 10.1186/1471-2105-3-17PubMed CentralView ArticlePubMedGoogle Scholar
- Durbin Blythe, Rocke David M.: Estimation of transformation parameters for microarray data. Bioinformatics 2003, 19: 1360–1367. 10.1093/bioinformatics/btg178View ArticlePubMedGoogle Scholar
- Schwarz Gideon: Estimating the dimension of a model. The Annals of Statistics 1978, 6: 461–464.View ArticleGoogle Scholar
- Eisen MB, Brown PO: DNA arrays for analysis of gene expression. Methods Enzymol 1999, 303: 179–205. 10.1016/S0076-6879(99)03014-1View ArticlePubMedGoogle Scholar
- Sokal Robert R., Rohlf F. James: Biometry. 3rd Edition New York, W. H. Freeman and Company 1995, 887.Google Scholar
- Metropolis Nicholas, Rosenbluth Arianna W., Rosenbluth Marshall N., Teller Augusta H., Teller Edward: Equation of state calculations by fast computing machines. Journal of Chemical Physics 1953, 21: 1087–1092.View ArticleGoogle Scholar
- Hastings WK: Monte Carlo sampling methods using Markov chains and their applications. Biometrika 1970, 57: 97–109.View ArticleGoogle Scholar
- Newton MA, Kendziorski CM, Richmond CS, Blattner FR, Tsui KW: On differential variability of expression ratios: improving statistical inference about gene expression changes from microarray data. Journal of Computational Biology 2001, 8: 37–52. 10.1089/106652701300099074View ArticlePubMedGoogle Scholar
- Gelman A, Roberts GO, Gilks WR: Efficient Metropolis jumping rules. Bayesian Statistics 5 (Edited by: Bernardo JM, Berger JO, Dawid AP and Smith AFM). Oxford University Press 1996, 599–607.Google Scholar
- Sudarsanam Priya, Iyer Vishwanath R., Brown Patrick O., Winston Fred: Whole-genome expression analysis of snf/swi mutants of Saccharomyces cerevisiae. Proceedings of the National Academy of Sciences - USA 2000, 97: 3364–3369. 10.1073/pnas.050407197View ArticleGoogle Scholar
- Alexandre H, Ansanay-Galeote V, Dequin S, Blondin B: Global gene expression during short-term ethanol stress in Saccharomyces cerevisiae. FEBS Letters 2001, 498: 98–103. 10.1016/S0014-5793(01)02503-0View ArticlePubMedGoogle Scholar
- Lyons Thomas J, Gasch Audrey P, Gaither L. Alex, Botstein David, Brown Patrick, Eide David: Genome-wide characterization of the Zap1p zinc-responsive regulon in yeast. Proceedings of the National Academy of the Sciences 2000, 97: 7957–7962. 10.1073/pnas.97.14.7957View ArticleGoogle Scholar
- Wernisch Lorenz: Can replication save noisy microarray data? Comparative and Functional Genomics 2002, 3: 372–374. 10.1002/cfg.196PubMed CentralView ArticlePubMedGoogle Scholar
- Pan Wei, Lin Jizhen, Le Chap T: How many replicates are required to detect gene expression changes in microarray experiments? A mixture model approach. Genome Biology 2002, 3: research0022.1. 10.1186/gb-2002-3-5-research0022Google Scholar
- Brazma Alvis, Hingamp Pascal, Quackenbush John, Sherlock Gavin, Spellman Paul, Stoeckert Chris, Aach John, Ansorge Wilhelm, Ball Catherine A., Causton Helen C., Gaasterland Terry, Glenisson Patrick, Holstege Frank C. P., Kim Irene F., Markowitz Victor, Matese John C., Parkinson Helen, Robinson Alan, Sarkans Ugis, Schulze-Kremer Steffen, Stewart Jason, Taylor Ronald, Vilo Jaak, Vingron Martin: Minimum information about a microarray experiment (MIAME)–toward standards for microarray data. Nature Genetics 2001, 29: 365–371. 10.1038/ng1201-365View ArticlePubMedGoogle Scholar
- Rockman Matthew V., Wray Gregory A.: Abundant raw material for cis-regulatory evolution in humans. Molecular Biology and Evolution 2002, 19: 1991–2004.View ArticlePubMedGoogle Scholar
- Yan Hai, Dobbie Zuzana, Gruber Stephen B., Markowitz Sanford, Romans Kathy, Giardiello Francis M., Kinzler Kenneth W., Vogelstein Bert: Small changes in expression affect predisposition to tumorigenesis. Nature Genetics 2002, 30: 25–26. 10.1038/ng799View ArticlePubMedGoogle Scholar
- Gibson G: Epistasis and pleiotropy as natural properties of transcriptional regulation. Theor Popul Biol 1996, 49: 58–89. 10.1006/tpbi.1996.0003View ArticlePubMedGoogle Scholar
- Doebley John, Lukens Lewis: Transcriptional regulators and the evolution of plant form. The Plant Cell 1998, 10: 1075–1082. 10.1105/tpc.10.7.1075PubMed CentralView ArticlePubMedGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media for any purpose, provided this notice is preserved along with the article's original URL.