Comparing transformation methods for DNA microarray data
 Helene H Thygesen^{1}Email author and
 Aeilko H Zwinderman^{1}
DOI: 10.1186/14712105577
© Thygesen and Zwinderman; licensee BioMed Central Ltd. 2004
Received: 02 February 2004
Accepted: 17 June 2004
Published: 17 June 2004
Abstract
Background
When DNA microarray data are used for gene clustering, genotype/phenotype correlation studies, or tissue classification the signal intensities are usually transformed and normalized in several steps in order to improve comparability and signal/noise ratio. These steps may include subtraction of an estimated background signal, subtracting the reference signal, smoothing (to account for nonlinear measurement effects), and more. Different authors use different approaches, and it is generally not clear to users which method they should prefer.
Results
We used the ratio between biological variance and measurement variance (which is an Flike statistic) as a quality measure for transformation methods, and we demonstrate a method for maximizing that variance ratio on real data. We explore a number of transformations issues, including BoxCox transformation, baseline shift, partial subtraction of the logreference signal and smoothing. It appears that the optimal choice of parameters for the transformation methods depends on the data. Further, the behavior of the variance ratio, under the null hypothesis of zero biological variance, appears to depend on the choice of parameters.
Conclusions
The use of replicates in microarray experiments is important. Adjustment for the nullhypothesis behavior of the variance ratio is critical to the selection of transformation method.
Background
When DNA microarray data are used for gene clustering or tissue classification, the signal intensities are usually transformed and normalized in several steps in order to improve comparability and signal/noise ratio. These steps may include subtraction of an estimated background signal, logarithmic transform, subtraction of the reference signal, and more. Different authors use different approaches: the most prominent controversies include Lowess smoothing, heteroscedastic models, subtraction of the reference signal and background correction. The aim of our study was to devise a method to find the optimal transformation method for a given dataset [1, 2].
As for Lowess smoothing, the idea is that the nonlinear trend, often found in the log (reference)/log (variety)scatter plot, is an artifact. Pritchard [3], who produced the data we used, expressed concern that this may not always be the case, and called for a method to distinguish artificial nonlinearity from biological effects. Also, at the CAMDA '02 Conference [4], concern was expressed that if the nonlinear effect is weak, the benefit from smoothing may not be worth its costs in term of possible overfitting.
As for heteroscedastic models, the idea is that the heteroscedasticity often found in the log (reference)/log (variety)scatter plot (the phenomena that the variance of the difference between the variety signal and the reference signal depends on the signal intensity), is an artifact. It follows that unbiased normalization would produce homoscedastic data. Some authors, including Rocke et al. [5], Huber et al. [6] and Cui et al. [7], suggest postnormalization homoscedasticity (termed variance stability) as a success criterion for normalization methods. This is also the basis of the heteroscedastic Lowess smoothing, adopted by SNOMAD [8]. However, unbiased normalization may not be synonymous with successful transformation from a practical point of view. To illustrate this, suppose that the true response (log(ratio)) is N(0, 1)distributed, whereas the noise is N(0, 1) for lowintensity genes and zero for highintensity genes. A variancestabilizing normalization would, then, shrink the expressions of the lowintensity genes with a factor , but for many applications, a lower shrinking factor (e.g., zero) would work better. (The National Institute of Aging online array analysis tool [9] solves this problem by estimating the technical variance locally). Actually, Brown et al. [10] showed how one can improve the discriminatory power of microarray data by explicitly reducing the weights of measurements suspected of being of low quality (shrinking). Although assigning explicit weights to measurements is uncommon, some methods do so implicitly. For example, using a linear scale instead of the log scale effectively shrinks response levels of lowexpressed genes. On the other hand, Huber et al. [6] found consistency between variance stabilization and reproducibility of posttransformation response.
Wernisch et al. [11] used the Pearson correlation coefficient for two replicates of the same experiment. Although that is probably a good success criterion in many cases, it cannot help us decide how to reduce the variety signal and the reference signal to one number, since the reference signal will typically show equally strong correlation as the variety signal. But the reference signal does not in itself tell us anything about the tissue sample.
He et al. [2] aimed at a success criterion more closely related to the application in question. They made the distinction between Type 1 applications (aimed at obtaining pvalues for the response of individual genes in parallel) and Type 2 applications (aimed at correlating multigene patterns to phenotypes). As for Type 1, they classified each gene as either differentially expressed or nondifferentially expressed (based on some threshold for the pvalue), and used the accuracy of the classification as a success criterion for the transformation. They proposed several methods for establishing the true set of differentially expressed genes.
Our choice of success criterion is similar to that of He [2], but instead of dichotomous judgements of differentiation, we use quantitative responses. The reasons for this choice are the following:

One application that we have in mind is the assessment of the normal variability of the expression of either individual genes or genes that are representative for some group of genes, e.g. genes involved in particular pathways (this was the purpose of Project Normal [3], for which the data we used were produced).

Even if the study is aimed at dichotomous judgements, the information loss from dichotomization may make the assessment of success less sensitive.

Using the quantitative response levels is simpler, because one does not need to decide on a threshold. Actually, Huber et al. [4], who used a success criterion similar to that of He [2], found that the performance of the Lowess smoother compared to the rank transform in some cases depends on the threshold.
Further, instead of comparing the posttransformation response estimate to true differentiation levels (which are generally unavailable), we measured posttransformation reproducibility in a replicated experiment. When more replicates of the same tissue sample are available, an obvious success criterion is a low withinsample variance. Consider an experiment in which S tissue samples are each hybridized independently on R arrays (replicates) with a probeset of G genes. Y_{ gsr }denotes some indicator (transformed signal) of the expression of gene g in replicate r of sample s. As a measure of reproducibility of the transformed signal, we use the the ratio of betweentissuesample variance to the withinsample variance (the Fstatistic), where the latter is averaged over the genes for stability:
Pvalue for the variance ratio
Part of the observed variance ratio can be explained as an artifact of statistical fluctuations. Assume, for simplicity, that Y_{ gsr }were N(0, σ)distributed (i.i.d.) for all g, s, r. Then, the expectation of the betweensample variance would become
VAR_{ between }≈ n_{ genes }σ^{2} (2)
whereas the expectation of the withinsample variance would become
VAR_{ within }≈ n_{ genes }n_{ repls }σ^{2} (3)
Since the number of genes is large, VAR_{ within }becomes reasonably stable, so the expected variance ratio would become
 1.
The pvalue for the variance ratio under the null hypothesis of zero biological variance.
 2.
The difference between the obtained variance ratio and the expected variance ratio (under the same null hypothesis).
In this paper, we show how those success criteria depend on parameter values for a number of transformation methods.
Results
Data
The study was based on the second data set for the CAMDA 2002 contest ("Project Normal") by Pritchard et al. [3]. In Project Normal, 18 tissue samples from 6 mice (from the testis, kidney and liver from each mouse) were hybridized independently on four identical microarrays. A mixture of all 18 samples was used as a reference. For each spot, the image analysis performed by Pritchard (using GenePix software [12]) provided four values (background and foreground for the variety and the reference). "Flagged" spots (related to absent genes or distorted parts of the image) were omitted. The background signal was ignored. Finally, all signal values were normalized with the median for the channel in question. We did not account for spatial effects, crosstalk or dye effects.
BoxCox transformation
The results suggested an optimal λvalue of 1 for the testis data, 1 for kidney and 10 for liver. These results were surprising, in particular the high variance ratios in the testis data set with negative λvalues, but the results became more plausible when compared to the expected variance ratios under the null hypothesis of zero biological variance.
We noticed that the variance of the permutation results became high for high λ values. This is not surprising: for high λ values, the dominance of a few, strongly expressed genes, becomes extreme, leading to less stable averages. We also noticed that the downward slope for negative λ values could be attributed to statistical fluctuations. Measured by pvalue, λ = 0.25 was the new optimum, but the differences within the range from 1 to 0.5 were hardly significant.
Partial subtraction of log(reference signal)
Because the quality of the reference signal is probably a property of the array rather than a biological effect, it may be too rigid to choose an overall subtraction factor for all 24 arrays for one tissue. In fact, for the second replicate of the testis of mouse 1, which we omitted because of bad image quality, ignoring the reference signal turned out to work better. As for partial subtraction of log(reference signal), our findings did not change when we corrected for the behavior of the variance ratio under the null hypothesis of zero biological variance.
Baseline shift
We computed variance ratios for the three data sets after shifting all measurements upwards by a number of medians (or averages) for the channel, and subsequently taking the logarithm. For the testis data, a baseline relative to the mean worked better than one relative to the median. For the kidney, the reverse was true. We also tried a baseline relative to the median (or average) background. The idea was that the baseline shift may work by reducing the influence from spots with a foreground signal not much stronger than the background. But a baseline relative to the foreground appeared to work much better.
Smoothing
The kidney data appeared to improve the most with a smooth band of 0.75, which is also the default in SPlus. The testis data did not improve through smoothing. We also tried to smooth log(variety) as a function of log (reference) (that is, without rotating the scatter plots), but that resulted in slightly lower variance ratios. Firstorder local regression (Lowess) generally worked slightly better than secondorder local regression (Loess). Also experiments that accounted for heteroscedasticity, by normalizing the residuals relative to the local standard deviation of the residuals, yielded disappointing results. That being said, the difference in variance ratio between the various smoothing methods was very small.
It is not clear why this difference between the testis data and the kidney data is observed. In general, the scatter plots for the testis data look nicer than those for the kidney data, suggesting that smoothing is not necessary for the testis data and only introduces an overfit which reduces the variance ratio. On the other hand, two of the mice had highly elevated expression levels of nearly all genes. Although the expression levels were divided by the global median, some tissuesamplespecific nonlinearity in those two samples may exist, e.g. because of a saturation effect. This would justify smoothing. However, smoothing would lead to a lower variance ratio because a tissuesample specific artifact can not be distinguished from a biological effect. This problem illustrates that independent replicates, where four different tissue samples from each mouse were used, would have been better. For the Lowess smoothing, adjusting for the behavior of the variance ratio under the null hypothesis did not change our findings.
Background correction
The background signal of a spot, estimated by the GenePix software, uses the median signal of all pixels in the neighbourhood of the spot. We investigated the effect of subtracting this background from the signals before transformation, which was either the rank transform, rank(variety)rank(reference), or log (variety)/log (reference). In the latter case, signals below some threshold were set equal to that threshold, and the procedure was repeated with different thresholds. We also tried using a kernel smoother, in which the local background estimates from Genepix were smoothed with an exponential kernel, restricted to the 12 nearest neighbour spots. In all cases, background correction slightly reduced the variance ratio.
Reproducibility of our findings
In a study such as this, attention should be paid to possible overfitting, in particular in the presence of outliers. The variance ratio is not a robust method, in principle, but since the variances are averages over 5776 genes, an outlier must be rather extreme if it is to skew the results significantly. Replacing the variance ratio with the slightly more robust measure
Discussion
The optimal transformation method will differ from application to application. Therefore, experiments should be designed in a way that makes it possible to identify the optimal transformation method. Replicating each array, as in the Project Normal datasets, is an excellent approach. Replicating each gene within each array would enable the same analysis.
The variance ratio is, as any performance measure would be, a stochastic entity. Therefore, its absolute value may be less meaningful than its pvalue under the hypothesis of zero biological variance. Under certain assumptions (normal distribution, homogene variance, zero correlation between genes), the pvalue of the variance ratio can be computed from the F distribution, but those assumptions were, in general, not met in this study. One can resort to a random permutation experiment in order to obtain that pvalue. This step is important, because optimizing the pvalue may lead to different parameter values than optimizing the crude variance ratio.
Conclusions
Background correction, based on the median of neighborhood pixel intensities, did not prove to be beneficial. It is possible that other image processing techniques could produce more useful background estimates.
As compared to the Standard Figures of Merit proposed by He et al. [2], our method is simpler, but He's method is more closely related to applications that rely on dichotomous judgements of differentiation of gene expressions. It will be interesting to see how the two approaches relate to each other if applied to the same data sets. Such a comparison may also resolve the question of whether the pergene response should be dichotomized or not.
Methods
General approach
For a number of different transformation methods, we computed the variance ratio for a number of parameter values, leading to a choice of parameter value that maximizes the variance ratio. This was done separately for each of three data sets (kidney, liver, testis), leading to three different optimal parameter values. Further, for a number of transformation methods, we computed the distribution of the variance ratio under the assumption of zero betweensample variance. The optimal parameter choice was then the one leading to the highest variance ratio, or the lowest pvalue, relative to the distribution of the variance ratio under the null hypothesis.
BoxCox transform
The common practice of taking the logarithm of the measured intensities can be justified by the assumption that most variance sources are multiplicative rather than absolute. Under that assumption, the logarithm of the intensity has desirable properties such as an approximately Gaussian distribution and homoscedastic residuals, i.e. all genes have equal influence on the model fitting, whether they give rise to high or low intensities. The question of whether to use a log scale or a linear one was raised by, among others, Efron et al. [13], who came to the conclusion that the log scale is better than the linear scale. If some variance sources (such as background noise) are absolute, one could use e.g., the square root as a compromise between a linear and a logarithmic scale, as in e.g., Amaratunga and Cabrera [14]. The generalization of this is the BoxCox transform
We used the following transform:
Y = B_{λ}(x_{ variety })  B_{λ}(x_{ reference }) (7)
for different values of λ, thereby finding the λvalue that led to the highest variance ratio. A consequence of this is that for a λvalue of 1 (linear scale), the reference signal is subtracted from the variety signal rather than divided by. We have chosen this deliberately. Alternatively, one could use
Y = B_{λ}(x_{ variety }/x_{ reference }).
Partial subtraction of log(reference signal)
It is sometimes argued, for example by Wolfinger [15], that one should ignore the reference signal. Although this partly depends on what biological question is asked (whether absolute gene expressions, or gene expressions relative to a given reference, are more meaningful), we will here find the degree to which the reference signal should be subtracted in order to maximize the variance ratio. The transformed signal becomes Y = log(x_{ variety })  a log(x_{ reference }) where a is a parameter. We have limited the analysis to the log transform, but the same analysis could easily be carried out with alternative methods, e.g., BoxCox.
Adding a baseline signal
When we restricted the analysis to genes in certain intensity ranges, it appeared that the log transform worked better for highintensity genes, and the linear scale better for lowintensity genes. This makes sense because the background signal (which is an additive variance source) is relatively more influential for lowintensity genes. So instead of the "static" BoxCox, which assumes the optimal λ to be the same for all genes, a "dynamic" BoxCox, with a λvalue approaching 1 for lowintensity genes and 0 for highintensity genes, may work better. A simple way to mimic this is to add some constant to all gene expressions before taking the logarithm. This additive constant, or baseline shift, could be set proportional to either the median or the average of the gene expressions for one channel – that is, we choose a different baseline shift for each channel. (Actually, because we normalized the data by setting the median of each channel to one, the baseline shift becomes the same for all channels if we choose a shift proportional to the median of the channel). The transformation becomes
Y = log(a x_{0,variety}+ x_{ variety })  log(a x_{0,reference}+ x_{ reference }) (8)
where x_{0} is either the average or the median expression over the genes, and a is a parameter. Other possible compromises between the linear scale and the logarithmic scale are the socalled Generalized Logarithm [5], and the arsinh transform [6].
Smoothing
Background correction
It can be debated whether background correction actually improves data quality. The reason why we have chosen not to correct for background is that many of the transformation methods that we studied cannot deal with negative numbers, which raises the question what to do with those spots that have higher background than foreground. A simple transformation, which does not have problems with negative values, is the rank. Thus, we applied the following transformations:
Y = rank(x_{ variety })  a rank(x_{ reference })
where a is a parameter. The variance ratios for these transformations were computed both with and without background correction. We also tried with the log transform, where signal values below a fixed threshold were set equal to that threshold. The arsinh transform suggested by Huber [6] is another solution to the problem with the negative signal intensities.
Pvalue for the variance ratio
As explained in section 1 (Background), the raw variance ratio may not be comparable for different transformation methods and different parameter values. Therefore, we computed the pvalue of that variance ratio relative to the null hypothesis that the expression of a gene depends only on the gene, and not on the tissue sample. The distribution of the variance ratio under that null hypothesis is computed with a simple permutation method, similar to the method described by Efron et al. [13]. The same permutation test is used by the National Institue of Aging [9], but they apply it locally to account for heteroscedasticity.
Consider the following design matrix for the unpermutated variance ratio (6 tissue samples (=mice), 4 replicates of each):
This means that the first replicate of the first tissue sample is substituted for the first replicate of the first sample, and so on.
In each permutated calculation, the rows were randomly permutated, e.g.,
Here, replicate 2 of the virtual tissue sample 1 refers to replicate 2 of physical tissue sample 4, and so on. Because we always substituted with a replicate with the same dye, dye effects are retained. On the basis of such a permutation, the variance ratio is computed. Repeating this with 500 different permutations yields 500 sample variance ratios. Finally, we compared the observed (unpermutated) variance ratio to the quantiles from the 500 samples.
Declarations
Authors’ Affiliations
References
 Tsodikov A, Szabo A, Jones D: Adjustments and measures of differential expression for microarray data. Bioinformatics 2002, 18: 251–260. 10.1093/bioinformatics/18.2.251View ArticlePubMedGoogle Scholar
 He YD, Dai H, Schadt EE, Cavet G, Edwards SW, Stepaniants SB, Duenwald S, Kleinhanz R, Jones AR, Shoemaker DD, Stoughton RB: Microarray Standard Data Set and Figures of Merit for Comparing Data Processing Methods and Experiment Designs. Bioinformatics 2003, 19(8):956–965. 10.1093/bioinformatics/btg126View ArticlePubMedGoogle Scholar
 Pritchard CC, Hsu L, Delrow J, Nelson PS: Project normal: Defining normal variance in mouse gene expression. PNAS 2001, 98(23):13266–71. 10.1073/pnas.221465998PubMed CentralView ArticlePubMedGoogle Scholar
 CAMDA '02 conference[http://www.camda.duke.edu]
 Rocke DM, Durbin B: Approximate variancestabilizing transformations for geneexpression microarray data. Bioinformatics 2003, 22(19(8)):966–72. 10.1093/bioinformatics/btg107View ArticleGoogle Scholar
 Huber W, Von Heydebreck A, Sultmann H, Poustka A, Vingron M: Variance stabilization applied to microarray data calibration and to the quantification of differential expression. Bioinformatics 2002, 18(Suppl 1):S96S104.View ArticlePubMedGoogle Scholar
 Cui X, Kerr MK, Churchill GA: Transformations for cDNA Microarray Data. Statistical Applications in Genetics and Molecular Biology 2003, 1(Article 4):2.Google Scholar
 Colantuoni C, Henry G, Zeger S, Pevsner J: SNOMAD (Standardization and Normalization of MicroArray Data): Webaccessible Gene Expression Data Analysis. Bioinformatics 2002, 18(11):1540–1. [http://pevsnerlab.kennedykrieger.org] 10.1093/bioinformatics/18.11.1540View ArticlePubMedGoogle Scholar
 National Institute of Aging: Online Microarray Data Analysis Tool.[http://lgsun.grc.nia.nih.gov/ANOVA]
 Brown CS, Goodwin PC, Sorger PK: Image metrics in the statistical analysis of DNA microarray data. PNAS 2001, 98(14):8944–49. 10.1073/pnas.161242998PubMed CentralView ArticlePubMedGoogle Scholar
 Wernisch L, Kendall SL, Soneji S, Wietzorrek A, Parish T, Hinds J, Butcher PD, Stoker NG: Analysis of wholegenome microarray replicates using mixed models. Bioinformatics 2003, 19(1):53–61. 10.1093/bioinformatics/19.1.53View ArticlePubMedGoogle Scholar
 Axon: GenePix Pro microarray scanning and analysis software.[http://www.axon.com/GN_GenePixSoftware.html]
 Efron B, Tibshirani R, Storey JD, Tusher V: Empirical Bayes Analysis of a Microarray Experiment. JASA 2001, 96(456):1151–1160. 10.1198/016214501753382129View ArticleGoogle Scholar
 Amaratunga D, Cabrera J: Analysis of Data From Viral DNA Microchips. JASA 2001, 96(456):1161–1170. 10.1198/016214501753381814View ArticleGoogle Scholar
 Wolfinger R, Chu T: Simultaneous Assessment of Transcriptomic Variability and Tissue Effects in Normal Mouse. Proceedings of CAMDA '02 2003. [http://www.camda.duke.edu]Google Scholar
 Venables WN, Ripley BD: Modern Applied Statistics with SPlus,. Springer Verlag 1994, 250.View ArticleGoogle Scholar
Copyright
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.