Sample size calculation for microarray experiments with blocked one-way design

Background One of the main objectives of microarray analysis is to identify differentially expressed genes for different types of cells or treatments. Many statistical methods have been proposed to assess the treatment effects in microarray experiments. Results In this paper, we consider discovery of the genes that are differentially expressed among K (> 2) treatments when each set of K arrays consists of a block. In this case, the array data among K treatments tend to be correlated because of block effect. We propose to use the blocked one-way ANOVA F-statistic to test if each gene is differentially expressed among K treatments. The marginal p-values are calculated using a permutation method accounting for the block effect, adjusting for the multiplicity of the testing procedure by controlling the false discovery rate (FDR). We propose a sample size calculation method for microarray experiments with a blocked one-way design. With FDR level and effect sizes of genes specified, our formula provides a sample size for a given number of true discoveries. Conclusion The calculated sample size is shown via simulations to provide an accurate number of true discoveries while controlling the FDR at the desired level.


Background
Clinical and translational medicine have benefited from genome-wide expression profiling across two or more independent samples, such as various diseased tissues compared to normal tissue. DNA microarray is a high throughput biotechnology designed to measure simultaneously the expression level of tens of thousands of genes in cells. Microarray studies provide the means to understand the mechanisms of disease. However, various sources of error can influence microarray results [1].
Microarrays also present unique statistical problems because the data are high dimensional and are insufficiently replicated in many instances. Methods of adjustment for multiple testing therefore become extremely important. Multiple testing methods controlling the false discovery rate (FDR) [2] have been popularly used because they are easy to calculate and less strict in controlling the false positivity compared to the family-wise error rate (FWER) control method [3].
Numerous sample size calculation methods have been proposed for comparing independent groups while controlling the FDR in designing microarray studies. Lee and Whitmore [4] considered comparing multiple groups using ANOVA models and derived the relationship between the effect sizes and the FDR using a Bayesian approach. Their power analysis does not address the multiple testing issue. Muller et al. [5] chose a pair of testing errors, including FDR, and minimized one while controlling the other at a specified level using a Bayesian decision rule. Jung [6] proposed a closed form sample size formula for a specified number of true rejections while controlling the FDR at a desired level. Pounds and Cheng [7] and Liu and Hwang [8] proposed similar sample size formulas which can be used for comparison of K independent samples. These methods are for the FDR-control methods based on independence or a weak dependency assumption among test statistics. Recently, Shao and Tseng [9] introduced an approach for calculating sample sizes for multiple comparisons accounting for dependency among test statistics.
In some studies, specimens for K treatments are collected from the same subject and means are compared across treatment groups. In this case, the gene expression data for the K treatments may be dependent since they share the same physiological conditions. For example, Feng et al. [10] conducted a study to discover the genes differentially expressed between center (C) and edge (E) of the uterine fibroid and the matched adjacent myometrium (M). In this study, specimens are taken from the three sites for each patient. The patients are blocks and the three sites (K = 3), C, E and M, are treatments (or groups) to be compared.
Since a set of K specimens are collected from each patient, we require a much smaller number of patients than a regular unblocked design. Furthermore, the observations within each block tend to be positively correlated, so that a blocked design requires a smaller number of arrays than the corresponding unblocked design just as a paired twosample design with a positive pairwise correlation requires a smaller number of observations than a two independent sample design. The more heterogeneous the blocks are, the greater the savings in number of arrays for the blocked design.
In this paper, we consider a non-parametric blocked F-test statistic to compare the gene expression level among K dependent groups. We adjust for multiple testing and control the FDR by employing a permutation method. We propose a sample size calculation method for a specified number of true rejections while controlling the FDR at a specified level. Through simulations, we show that the blocked F-test accurately controls the FDR using the permutation resampling method and the calculated sample size provides an accurate number of true rejections while controlling the FDR at the desired level. For illustration, the proposed methods are applied to the fibroid study [10] mentioned above.

Non-parametric block F-test statistic
Suppose that we want to discover genes that are differentially expressed among K sites (treatments or groups). For each of n patients (blocks), a specimen is collected from each site for a microarray experiment on m genes. In this case, the gene expression data from the K sites tend to be correlated. Let Y ijk denote the expression level of gene i (= 1,..., m) from treatment k (= 1,..., K) of block j (= 1,..., n). We consider the blocked one-way ANOVA model where, for gene i, μ i is the population mean, δ ik is a fixed treatment effect and the primary interest, γ ij is a random block effect, and ε ijk is a random error term. We assume that , γ i1 ,..., γ in are independent and identically distributed (IID) with mean 0 and variance v i , (ε ijk , 1 ≤ j ≤ n, 1 ≤ k ≤ K) are IID with mean 0 and variance , and error terms and block effects are independent. The standard ANOVA theory using parametric F distributions to test the treatment effect assumes a normal distribution for ε ijk . However, in this paper, we avoid the normality assumption by using a permutation resampling method in testing and a large-sample approximation in sample size calculation.
For gene i(= 1,..., m), the hypotheses for testing the treatment effect are described as   [11] can be used to implement the permutation-based multiple testing procedure for blocked microarray data. We consider adjusting for the multiplicity of the testing procedure by controlling the FDR [12,13]. (iv) For a chosen constant λ ∈ (0, 1), estimate the q-value by (v) For a specified FDR level q*, discover gene i (or reject H i ) if q i < q*. and ν 2 degrees of freedom, and noncentrality parameter η.
The marginal powers are expressed as where denotes the 100(1 -α) percentile of distribution. The marginal powers can be calculated using R, SAS or some other packages. Suppose we want r 1 true rejections while controlling the FDR at q*. By combining this with (3) and (4), we obtain two equations and Note that r 1 /m 1 denotes the probability of true rejection.
At the design stage of a study, m is given by the number of genes included in the chips to be used for microarray experiment, m 1 and { , i ∈ 1 , 1 ≤ k ≤ K} are projected based on biological knowledge or estimated from pilot data, and K, r 1 (or r 1 /m 1 ) and q* are prespecified. The only unknown variables in (5) and (6) are α and n. By solving (6) with respect to α, we obtain α* = r 1 q*/{m 0 (1 -q*)} and, by plugging this in (5), we obtain an equation for r 1 depending only on n, The marginal power function (4) includes n in the degrees of freedom of the denominator as well as the noncentrality parameter of the F-distributions. The impact of the degrees of freedom of the denominator of the F-statistic on the marginal power is much weaker than that of the noncentrality parameter, so that β i (α) is a monotone increasing function of n, and consequently equation (7) has a unique solution. Figure  In summary, the sample size (i.e., number of blocks) n for r 1 (≤ m 1 ) true rejections is calculated as follows, assuming that the error terms in model (1) are normally distributed.
Sample size calculation based on the noncentral F-distribution (i) Specify the input variables: -K = number of treatments; -m = total number of genes for testing; d ik (iii) The required sample size is n blocks, or nK array chips.
In the sample size formula based on the noncentral F-distribution, the relationship between n and the marginal power functions based on the F-distribution is complicated and a normal distribution assumption of the error terms is required. In the large sample case, we can loosen the normality assumption and simplify this relationship.
If the error terms have a finite 4-th moment, then, for large n, the distribution of F i is approximated by A proof is given in the Appendix. Similarly, for large n, the is the 100(1 -α) percentile of the χ 2 distribution with ν degrees of freedom. Hence, the marginal power for Equation (8) is especially useful when we want to compare the powers between a blocked one-way design and an unblocked one-way design. Using similar approximations, it is easy to show that an approximate sample size N = nK for a study with unblocked one-way design with a balanced allocation is obtained by solving with respect to n, where . The only difference between (8) and (9) (8) and (9), we have n u = (1 As an example, consider the design of the fibrosis study as discussed in Background Section and suppose that the variance of the block effects is half of that of measurement errors for the prognostic genes, i.e. f = 0.5. In this case, if a blocked design requires n b = 100 patients and 3n b = 300 array chips, then the corresponding unblocked design with a balanced allocation requires n u = 150 patients per group or a total 450 patients. For an unblocked design, the number of array chips is identical to that of patients, and compared to the blocked design, the unblocked design requires 1.5 times more chips and 4.5 times more patients.

Results and discussion
Simulations First, we investigate the accuracy of the FDR control based on blocked one-way ANOVA tests and the sample size formulas via simulations. For the simulations on FDR control, we consider blocked one-way designs with K = 3 treatments and n = 10,  l(= 1,..., N), the FDR-control multiple testing method is applied to the simulated data using tuning parameter λ = 0.95 [12] to count the numbers of total rejections and false rejections and to estimate the FDR, .
Then the empirical FDR is obtained as   Table 2 summarizes the simulation results by the two methods. As expected, sample size increases in r 1 and decreases in m 1 and q*. Since the standardized effect sizes for the differentially expressed genes influence the sample size through their sum of squares, the combination of effect sizes (1/4, 0, -1/4) requires a larger sample size than (1/4, -1/2, 1/4). The sample size based on the chi-square approximation is always smaller than that based on the Fdistribution. The median (Q 2 ) of the empirical true rejections is smaller than the nominal r 1 for the sample size based on the chi-square approximation, especially with a small n, while the sample size based on the F-distribution is always accurately powered, i.e. Q 2 ≈ r 1 .

Example
We applied the permutation-based blocked one-way ANOVA and the sample size calculation method to the fibroid study discussed in the Background Section. From each patient, specimens are taken from two sites of fibroid tissue, center (C) and edge (E), and one normal myometrium (M). Five patients are accrued to the study. We regard the three sites as treatments (K = 3) and the patients as blocks (n = 5). mRNA was amplified and hybridized onto HG-U133 GeneChips according to the protocols recommended by Affymetrix (Santa Clara, CA), and m = 54675 probe sets on the array were analyzed. Expression values were calculated using the Robust Multichip Average (RMA) method [14]. RMA estimates are based upon a robust average of background corrected PM intensities. Normalization was done using quantile normalization [15]. We filtered out all "AFFX" genes and genes for which there were 4 or fewer present calls (based on Affymetrix's present/marginal/absent (PMA) calls using mismatch probe intensity, the ratio of PM to MM). That is, a gene is included only if there are at least 3 present calls among the 15 PMA calls. Filtering yielded 30711 genes to be used in the subsequent analyses.
In order to group the samples according to the degree of similarity present in the gene expression data, we first applied a hierarchical clustering analysis to the filtered 30711 gene expression data and generated a dendrogram Based on the chi-square approximation Figure 2). We used the Complete Linkage method [16] and Pearson's correlation coefficient as a measure of similarity. In the dendrogram, the height of each branch point indicates the similarity level at which each cluster was generated. We obtained the same clustering using the L 2 norm as a measure of similarity. Except for patient 2, E and C are clustered together for each patient. In spite of the block effect, M is clustered separately from E and C regardless of patient assignment. We conclude that C and E have similar gene expression profiles, but M has a different gene expression profile from either C or E. While the clustering analysis investigates the genome wide expression profile, blocked one-way ANOVA helps us identify individual genes differentially expressed among the three sites. Using the blocked one-way ANOVA method, we selected the top 50 genes in terms of parametric p-values ( Table 3). The expression patterns of six genes that are identified as differentially expressed are presented in Figure 3. The expression levels of each patients are connected among three sites. These genes are similarly expressed between C and E, but differentially expressed in M. Further, 220273_at, 210255_at, 229160_at, 204620_s_at and 217287_s_at are under-expressed in M while 1553194_at is over-expressed in M.
The results of our analysis of the two sites of fibroid tissue, center and edge, compared to the normal myometrium using a blocked one-way design suggest that reduced FDR provides an enhanced approach to clinical microarray studies. Our findings are consistent with previously reported genome-wide profiling studies [17,18]. We believe that these results support the hypothesis that uterine fibroids develop through altered wound healing signaling pathways leading to tissue fibrosis [19,20]. Using the method described in this paper, genes differentially over-expressed in the fibroid tissue compared to myometrium are related to extracellular matrix (ECM) and ECM regulation such as collagen IV, alpha 1, versican (chondroitin sulfated 2) and IL-17β [21]. IL-17β, a cellcell signaling transducer has been reported to enhance MMP secretion and to rapidly induce phosphorylation of the extracellular signal-related kinases (ERK) 1/2 and p38MAPK in colonic myofibroblasts and has been shown to stimulate MMP-1 expression in cardiac fibroblasts through ERK 1/2 and p38 MAPK [22,23]. Thus IL-17β is important in remodelling of the extracellular matrix. According to our analysis, RAD51-like 1, a recombinational repair gene, is also over-expressed in fibroids, which is consistent with a report that RAD51B is the preferential translocation partner of high mobility group protein gene (HMGIC) in uterine leiomyomas [24]. HMGIC codes for a protein that is a non-histone DNA binding factor that is expressed during development in embryonic tissue and is an important regulator of cell growth, differentiation and transformation as well as apoptosis [25]. Arrest of apoptosis appears to be a hallmark of uterine fibroids, a finding that is characteristic of altered wound healing as well [19]. HMGIC appears to play a role in the development of uterine fibroids [19,26,27].
Suppose that we want to design a new fibroid study using the data analyzed above as pilot data. In the sample size calculation, we set m = 30, 000. We assume that the m 1 = 50 genes which were selected as the top 50 genes in terms of parametric p-value are differentially expressed in the three sites (K = 3). From the pilot data, we estimate the standardized treatment effect δ ik . For illustration, the effect sizes of these m 1 = 50 genes are taken to be δ ik = 0.1 . We need n = 15 patients (blocks) to discover 90% of the prognostic genes, i.e. r 1 = [0.9 × 50] = 45, while controlling the FDR at q* = 5% level. In a simulation study, we generated N = 1000 microarray data sets of size n = 15 under this design setting. With q* = 0.05, we observed the quartiles Q 2 (Q 1 , Q 3 ) = 46(45, 47) from the empirical distribution of the observed true rejections.  Expression patterns of six genes that are significantly differentially expressed in three sites Figure 3 Expression patterns of six genes that are significantly differentially expressed in three sites.

Conclusion
We have considered studies where microarray data for K treatment groups are collected from the same subjects (blocks). We discover the genes differentially expressed among K groups using non-parametric F-statistics for blocked one-way ANOVA while controlling the FDR. We employ a permutation method to generate the null distribution of the F-statistics without a normal distribution assumption for the gene expression data. The permutation-based multiple testing procedure can be easily modified for controlling the familywise error rate, see e.g. Westfall and Young [28] and Jung et al. [29].
We propose a simple sample size calculation method to estimate the required number of subjects (blocks) given the total number of genes m, number of differentially expressed genes m 1 and their standardized effect sizes ( , 1 ≤ i ≤ m 1 , 1 ≤ k ≤ K) and the number of true rejections r 1 at a specified FDR level q*. Through simulations and analysis of a real data set, we found that the permutation-based analysis method controls the FDR accurately and the sample size formula performs accurately. While we specify the individual effect sizes for the prognostic genes, some investigators [30,31] use a mixture model for the marginal p-values by specifying a distribution for the effect sizes among m genes.
Glueck et al. [32] propose an exact calculation of average power for the Benjamini-Hochberg [2] procedure for controlling the FDR. Their formula may is useful for deriving sample sizes when the test statistics are independent and the number of hypotheses m is small. However, it is not appropriate for designing a microarray study with a large number of dependent test statistics.

Appendix
We want to prove that F i converges to in distribution regardless of the normal distribution assumption on ijk and γ ij . We only assume that . The following is one of key lemmas used to derive the distribution of the F-statistics in the standard ANOVA theory, see e.g. Section 3b.4 of Rao [33].
We research and participated in the writing of the manuscripted. PCL was responsible for the study design, conduct and oversight of the experiments and interpretation of results. She contributed to the preparation of the manuscript. LF was responsible for preparing the tissue samples for microarray analysis and interpretation of results and in manuscript preparation. The authors are solely responsible for the content of this study. All authors read and approved the final manuscript.