Multiple testing for gene sets from microarray experiments
 Insuk Sohn^{1},
 Kouros Owzar^{2},
 Johan Lim^{3},
 Stephen L George^{2},
 Stephanie Mackey Cushman^{4} and
 SinHo Jung^{2}Email author
https://doi.org/10.1186/1471210512209
© Sohn et al; licensee BioMed Central Ltd. 2011
Received: 13 December 2010
Accepted: 26 May 2011
Published: 26 May 2011
Abstract
Background
A key objective in many microarray association studies is the identification of individual genes associated with clinical outcome. It is often of additional interest to identify sets of genes, known a priori to have similar biologic function, associated with the outcome.
Results
In this paper, we propose a general permutationbased framework for gene set testing that controls the false discovery rate (FDR) while accounting for the dependency among the genes within and across each gene set. The application of the proposed method is demonstrated using three public microarray data sets. The performance of our proposed method is contrasted to two other existing Gene Set Enrichment Analysis (GSEA) and Gene Set Analysis (GSA) methods.
Conclusions
Our simulations show that the proposed method controls the FDR at the desired level. Through simulations and case studies, we observe that our method performs better than GSEA and GSA, especially when the number of prognostic gene sets is large.
Keywords
Background
One of the primary objectives in microarray association studies is the identification of individual genes that are associated with clinical endpoints such as disease type, toxicity or time to death. It is also of interest to examine the association between known biological categories or pathways and outcome. To this end, gene sets a priori believed to have similar biological functions from databases including KEGG [1] and Gene Ontology [2] are used. In recent years, a number of statistical methods have been proposed for the identification of significant genesets based on microarray experiments. Ackerman and Strimmer [3] list 36 methods, including [4–13], while outlining a general framework for formulating the hypothesis and analysis method for gene set inference.
In this paper, we propose a gene set analysis framework that utilizes classical theory for estimating equations to assess the association between each gene set and the outcome of interest. One of the statistical challenges in this setting is that there is dependency within each gene set, by virtue of coregulated genes belonging to the same gene set, as well as dependency across the gene sets since gene sets are not mutually exclusive. Our method will account for both intragene set and intergene set dependencies. Furthermore, given the large number of gene sets, one has to address the issue of multiple testing. The sampling distribution of our proposed procedure is approximated using permutation resampling to simultaneously address the dependency and multiple testing issues by controlling the false discovery rate (FDR; [14]). In the framework described by Ackerman and Strimmer [3], gene set analysis methods are broadly categorized as univariate or as global and multivariate procedures. Generally speaking, our method belongs to the latter category. The novelty of our proposed approach is that it leverages the flexibility of estimating equations to conduct inference for a variety of endpoints including binary, continuous, censored or longitudinal outcomes.
After presenting the theoretical and computational details for the proposed method, we summarize the results from a simulation study evaluating its statistical properties. We then apply the proposed method to analyze a number of microarray data sets. Finally, we provide a brief discussion to compare the performance of our method to those of two other methods: GSEA [6] and GSA [7]. For notational brevity, we will refer to transcripts on microarrays as genes, even though this may not be technically correct.
All analyses are carried out using the R statistical environment [15]. The code is available from http://www.duke.edu/~is29/GeneSet. Generalized inverses are computed using the pinv function from the maanova[16] extension package. The inverse of linear shrinkage covariance matrix is computed using invcov.shrink function in corpcor [17] extension package. The R extension packages RGSEA[6] and GSA[18] are used to implement the GSEA and GSA methods respectively. The qvalue[19] extension package is used for calculating FDR adjusted Pvalues. For gene set and probe set annotation, Bioconductor [20] annotation packages (e.g., hu6800.db[21]) and Molecular Signature Database (MSigD; http://www.broad.mit.edu/gsea) annotation files are used.
Methods
In these discussions, we will assume that RNA expression levels for m genes have been measured for n patients. Let us denote the set of genes on the microarray by G = {G_{1}, ..., G_{ m } }. For patient i(= 1, ..., n), let y_{ i } denote the clinical outcome and z_{ ij } denote the measured gene expression level for G_{ j } . Let G_{ j } ⊥ Y denote that expression of gene j is not associated with outcome. For each gene the marginal inference of interest will be canonically presented as testing H_{ j } : G_{ j } ⊥ Y versus .
where U_{ ij } (θ_{ j } ) is a function of the data for subject i only so that U_{1j}, ...,U_{ nj } are independent. The corresponding test statistic for H_{ j } will be U_{ j } (0). Let μ_{ ij } (θ_{ j } ) = E(U_{ ij } ) and . If E{U_{ j } (θ)} is a smooth function and E{U_{ j } (θ)} = 0 has a unique solution, then the solution to U_{ j } (θ) = 0 is a consistent estimator of θ_{ j } . The family of score statistics [22] is a special case of this type of estimating equation.
A gene set is defined as a subset of G. We will assume that there are K prespecified gene sets say based on a given annotation database such as KEGG or GO. Note that is usually a proper subset of G as not all genes are annotated. Let m_{ k } (k = 1, ..., K) denote the number of genes in gene set . We consider a gene set to be associated with the outcome of interest if at least one of its member genes is associated with the outcome. Let denote that gene set is not associated with the outcome Y. The hypotheses of interest from gene set k can then be denoted as testing versus .
respectively, where . Let and .
where . If n is large and m_{1} < n, the distribution of W_{1} under is approximately χ^{2} with m_{1} degrees of freedom. Similarly, we can compute U_{ k }, V_{ k } and W_{ k } for any gene set .
In many cases, the sample size for a microarray study may not be large enough for the null sampling distribution to be well approximated by the theoretical limiting distribution. To address this issue, we propose calculating the Pvalues by approximating the exact null sampling distribution using permutation resampling. Note that the permutation distribution is generated under the hypothesis . That is, none of the K gene sets are associated with the outcome. This hypothesis is equivalent to the hypothesis ∩ _{ j }H_{ j } (i.e., none of the genes are associated with outcome). Note that the latter intersection is restricted to G*, the set of annotated genes. A permutation replicate sample is obtained by randomly shuffling the the clinical outcomes {y_{1}, ..., y_{ n } } while holding the gene expression matrix in place. This ensures that the intragene dependency structure is preserved while breaking the association between the genes and the outcome.
If m_{ k } is large, V_{ k } many not be reliably inverted numerically and, in the case where it exceeds n, is not invertible. For these cases, we consider the MoorePenrose (MP) generalized inverse or the inverse of the linear shrinkage covariance matrix estimate V_{LW}[13, 23–25]. Here, we remark that the Hotelling's tests with the MP generalized inverse (ginverse) and that with the inverse of V_{LW} have been previously studied [13, 23]. The MP ginverse (of the sample covariance matrix) uses to derive a test statistic , where P_{ k } is the eigen matrix and , ν_{1}, ..., ν_{ d } are the d positive eigenvalues of V_{ k } . The asymptotic distribution of W_{ k } when m_{ k } is larger than n has been investigated extensively (e.g., [26–28]). The linear shrinkage estimate (LW) of V is V_{LW} = λV + (1  λ)E, where E is a well conditioned target matrix and λ is the tuning parameter. The tuning parameter λ is chosen to minimize the Frobenius risk along with several candidates of target matrices [24, 25].
TwoSample Tests
where is the pooled variancecovariance matrix. For θ_{ j } = E(z_{1ij})  E(z_{2ij}) and μ_{ ij } (θ_{ j } ) = θ_{ j } , T^{2} asymptotically has a distribution under H_{0}.
where and . Since T^{2} is asymptotically equivalent to W = U^{ T }V^{1}U under H_{0}, we use the more popular Hotelling's T^{2} statistic in this paper.
As a rank test alternative to the ttest, it is easy to show that the Wilcoxon rank sum test can be expressed as T^{2} with z_{ gij } the rank of the gene j expression level for subject i in the pooled data {z_{ gij } , 1 ≤ i ≤ n_{ g } , g = 1, 2}. In this case, θ_{ j } = P(z_{1ij}≥ z_{2ij})  1/2 and μ_{ ij } (θ_{ j } ) = θ_{ j } .
Linear Regression Case
Cox Regression Case
The resulting variance estimator is equivalent to the robust estimator under the possible violation of the proportional hazards model proposed by Lin and Wei [30].
Results
Simulation Study
where s_{ i } = 0 if subject i belongs to group 1 and s_{ i } = 1 otherwise, and ρ_{3} = 1  ρ_{1}  ρ_{2}. For gene j, δ_{ j } is the treatment effect, D is the number of prognostic genes, K_{1} is the number of prognostic gene sets, a_{ k } is the gene set effect, b is the array effect, ρ_{1} and ρ_{2} are the within gene sets and within arrays correlation coefficients resepctively, and ε_{ ijk } is the error term. The gene set effect a_{ k } , the array effect b, and the error term ε_{ ijk } are generated from independently and identically distributed N(0, 1) random variate.
Empirical FDR and mean true rejections
q* = 0.01  q* = 0.05  q* = 0.1  q* = 0.2  

( m _{ k } , K )  K _{ 1 }  D / m  ( ρ _{ 1 } , ρ _{ 2 } )  
(50,20)  1  0.2  (0, 0)  0.010  0.10  0.063  0.25  0.105  0.35  0.189  0.46 
(0.2, 0.2)  0.015  0.08  0.053  0.28  0.128  0.38  0.231  0.45  
(0.4, 0.4)  0.013  0.12  0.045  0.25  0.087  0.36  0.215  0.47  
0.5  (0, 0)  0.011  0.72  0.060  0.88  0.107  0.95  0.220  0.98  
(0.2, 0.2)  0.015  0.71  0.051  0.89  0.122  0.94  0.212  0.97  
(0.4, 0.4)  0.005  0.74  0.071  0.89  0.115  0.94  0.266  0.95  
0.8  (0, 0)  0.013  0.97  0.057  1.00  0.106  1.00  0.215  1.00  
(0.2, 0.2)  0.013  0.98  0.065  1.00  0.138  1.00  0.235  1.00  
(0.4, 0.4)  0.018  0.97  0.078  1.00  0.124  1.00  0.253  1.00  
5  0.2  (0, 0)  0.025  0.72  0.067  1.70  0.136  2.51  0.239  3.56  
(0.2, 0.2)  0.009  0.68  0.056  1.79  0.127  2.63  0.235  3.59  
(0.4, 0.4)  0.011  0.74  0.058  1.82  0.118  2.46  0.232  3.55  
0.5  (0, 0)  0.007  4.44  0.058  4.94  0.124  4.97  0.227  5.00  
(0.2, 0.2)  0.013  4.46  0.056  4.90  0.118  4.97  0.214  4.99  
(0.4, 0.4)  0.012  4.44  0.074  4.87  0.137  4.96  0.239  5.00  
0.8  (0, 0)  0.012  4.99  0.066  5.00  0.120  5.00  0.235  5.00  
(0.2, 0.2)  0.011  5.00  0.056  5.00  0.105  5.00  0.200  5.00  
(0.4, 0.4)  0.011  4.98  0.053  5.00  0.101  5.00  0.218  5.00  
(20,50)  1  0.2  (0, 0)  0.008  0.04  0.055  0.13  0.092  0.21  0.185  0.25 
(0.2, 0.2)  0.015  0.04  0.051  0.10  0.128  0.15  0.219  0.21  
(0.4, 0.4)  0.005  0.06  0.050  0.10  0.110  0.14  0.198  0.22  
0.5  (0, 0)  0.013  0.45  0.048  0.63  0.118  0.72  0.240  0.82  
(0.2, 0.2)  0.010  0.43  0.077  0.64  0.122  0.72  0.223  0.80  
(0.4, 0.4)  0.010  0.49  0.072  0.68  0.112  0.76  0.214  0.83  
0.8  (0, 0)  0.010  0.86  0.043  0.97  0.113  0.99  0.222  0.99  
(0.2, 0.2)  0.015  0.89  0.048  0.98  0.115  0.98  0.208  0.98  
(0.4, 0.4)  0.013  0.86  0.052  0.96  0.102  0.99  0.201  1.00  
5  0.2  (0, 0)  0.013  0.31  0.054  0.76  0.121  1.08  0.210  1.64  
(0.2, 0.2)  0.010  0.28  0.039  0.57  0.102  0.89  0.224  1.56  
(0.4, 0.4)  0.015  0.18  0.062  0.57  0.103  0.94  0.195  1.59  
0.5  (0, 0)  0.011  3.03  0.055  4.03  0.107  4.43  0.201  4.75  
(0.2, 0.2)  0.008  3.01  0.054  4.11  0.104  4.44  0.218  4.77  
(0.4, 0.4)  0.016  3.22  0.058  4.16  0.103  4.50  0.203  4.72  
0.8  (0, 0)  0.010  4.74  0.054  4.91  0.112  4.95  0.224  4.99  
(0.2, 0.2)  0.011  4.76  0.054  4.94  0.111  4.97  0.212  4.98  
(0.4, 0.4)  0.012  4.73  0.054  4.93  0.110  4.96  0.201  4.99 
Empirical FDR and mean true rejections on simulation data with small n large p values.
q* = 0.01  q* = 0.05  q* = 0.1  q* = 0.2  

K _{ 1 }  D / m _{ k }  ( ρ _{ 1 } , ρ _{ 2 } )  method  
1  0.2  (0, 0)  MP  0.0160  0.004  0.0633  0.018  0.1102  0.036  0.2328  0.058 
LW  0.0060  0.018  0.523  0.040  0.1022  0.060  0.2294  0.098  
(0.2 0.2)  MP  0.0100  0.006  0.0593  0.010  0.1145  0.022  0.2325  0.042  
LW  0.0180  0.006  0.0563  0.020  0.1058  0.036  0.2249  0.084  
(0.4 0.4)  MP  0.0180  0.008  0.0590  0.020  0.1133  0.032  0.2182  0.054  
LW  0.0100  0.010  0.0503  0.032  0.1087  0.048  0.2387  0.094  
0.5  (0, 0)  MP  0.0120  0.024  0.0650  0.068  0.1138  0.100  0.2016  0.162  
LW  0.0210  0.078  0.0652  0.160  0.1189  0.230  0.2303  0.330  
(0.2 0.2)  MP  0.0300  0.024  0.0720  0.072  0.1269  0.106  0.2608  0.170  
LW  0.0200  0.084  0.0823  0.154  0.1467  0.222  0.2658  0.316  
(0.4 0.4)  MP  0.0290  0.038  0.0920  0.072  0.1359  0.112  0.2161  0.184  
LW  0.0150  0.070  0.0747  0.168  0.1300  0.228  0.2444  0.338  
0.8  (0, 0)  MP  0.0150  0.084  0.0667  0.136  0.1261  0.178  0.2534  0.258  
LW  0.0070  0.240  0.0520  0.404  0.1199  0.508  0.2179  0.618  
(0.2 0.2)  MP  0.0260  0.068  0.0563  0.142  0.1147  0.198  0.2341  0.276  
LW  0.0220  0.234  0.0603  0.420  0.1054  0.516  0.2095  0.622  
(0.4 0.4)  MP  0.0153  0.068  0.0738  0.138  0.1299  0.176  0.2548  0.258  
LW  0.0270  0.274  0.0607  0.434  0.1129  0.512  0.2118  0.636  
5  0.2  (0, 0)  MP  0.0080  0.022  0.0503  0.094  0.0924  0.194  0.1817  0.386 
LW  0.0120  0.066  0.0443  0.188  0.0933  0.344  0.1883  0.726  
(0.2 0.2)  MP  0.0180  0.024  0.0683  0.084  0.1030  0.138  0.2140  0.364  
LW  0.0170  0.054  0.0642  0.144  0.1028  0.288  0.2125  0.646  
(0.4 0.4)  MP  0.0060  0.040  0.0350  0.078  0.0848  0.158  0.1969  0.304  
LW  0.0150  0.072  0.0446  0.192  0.0762  0.328  0.1761  0.668  
0.5  (0, 0)  MP  0.0143  0.158  0.0479  0.402  0.0990  0.608  0.2144  1.062  
LW  0.0140  0.442  0.0626  1.148  0.1221  1.734  0.2131  2.586  
(0.2 0.2)  MP  0.0157  0.148  0.0586  0.418  0.0974  0.662  0.2090  1.128  
LW  0.0127  0.452  0.0646  1.162  0.1076  1.796  0.2165  2.644  
(0.4 0.4)  MP  0.0160  0.150  0.0602  0.418  0.0978  0.678  0.2062  1.142  
LW  0.0108  0.498  0.0572  1.154  0.1134  1.680  0.2226  2.548  
0.8  (0, 0)  MP  0.0193  0.468  0.0662  1.006  0.1194  1.452  0.2320  2.054  
LW  0.0216  1.660  0.0609  3.044  0.1220  3.702  0.2385  4.286  
(0.2 0.2)  MP  0.0127  0.478  0.0567  1.036  0.1130  1.440  0.2257  2.156  
LW  0.0145  1.654  0.0506  2.970  0.1157  3.670  0.2269  4.278  
(0.4 0.4)  MP  0.0200  0.510  0.0587  1.020  0.1080  1.468  0.2163  2.164  
LW  0.0201  1.772  0.0729  3.066  0.1303  3.662  0.2350  4.234 
We compare the performance of our method to GSEA and GSA within the simulation framework described above. We choose, for GSEA, the weighted Kolmogorov Smirnovlike statistic as enrichment correlationbased weighting, while for GSA we choose the maxmean statistic along with restandardization. The technical details are provided in [6] and in [7] respectively.
We generate m = 1, 000 genes and n = 100 samples, each with nonoverlapping K = 50 gene sets of m_{ k } = 20 genes, (ρ_{1}, ρ_{2}) = (0, 0), D/m_{ k } = 1, and δ = 0.4 as in [7]. The first (n_{1} = 50) and second (n_{2} = 50) samples will constitute the control and treatment groups respectively. Next, we will discuss two scenarios similar to those considered by [7]:

Onesided shifts: The mean expression level for the m_{ k }= 20 genes in each of the K_{1} prognostic gene sets is δ = 0.4 units higher in the treatment group.

Twosided shifts: The mean expression level for the first 10 genes in each of the K_{1} prognostic gene sets is δ = 0.4 units higher, while the mean expression level for the next 10 genes is δ = 0.4 units lower.
Case Studies
TwoSample Case
Cox Regression Case
Discussion
For the Gender data set, at the FDR level of q* = 0.2, our method identifies 8 gene sets compared to only 4 for the other two methods [see Additional file 1]. There are 4 prognostic gene sets identified in common among the three methods, consisting of gene sets found on ChrY, ChrYp11, ChrYq11, and ChrXp22. Our method identifies 4 other gene sets not identified by the other two methods, which include gene sets for ChrX, ChrXp11, Chr3q25, and Chr6q25. Genes expressed on the Y chromosome are expected to be differentially expressed between genders, while gene expression from the X chromosome is more similar between genders due to X chromosome inactivation in females [33, 34]. However, ChrXp22 and ChrXp11 gene sets have been previously been shown to be overrepresented in females likely caused by escape of X inactivation [35]. Furthermore, several genes within the Chr3q25 and Chr6q25 gene sets have also been shown to be differentially expressed between males and females, including ACAT2 [36], MAP3K4 [37], NOX, PTX3 [38], SGEF, and SOD2 [39]. Thus, our method for identifying overrepresented genes in gene set lists can provide biologically relevant and important information that may be overlooked by other common methods such as GSA and GSEA.
For the p53 data set, at the same FDR level, our method identifies 87 prognostic gene sets while GSA and GSEA identify 5 and 9 prognostic gene sets, respectively [see Additional file 1]. There are 5 prognostic gene sets common among the three methods, including the p53 pathway, hsp27 pathway, radiation sensitivity pathway, ceramide pathway, and the ras pathway. However, our method identifies 78 gene sets not identified by the other two methods. Additional file 1 also provides a list of gene sets that are identified only by our method. p53 is a tumor suppressor protein that is activated in response to DNA damage. p53 can induce growth arrest by halting the cell cycle at the G1/S phase transition to allow DNA repair or it can induce apoptosis if the DNA damage cannot be repaired. p53 acts as a transcription factor regulating the expression of many genes involved in its functions [40]. Thus many of the gene sets identified by our method can be directly linked to p53 functions, such as cell cycle arrest, ATM pathway, tumor suppressor, bcl2 family and network, death pathway, etc [40]. Additionally, several cytokine and growth factor signaling pathways are represented in our list of gene sets differentially expressed between p53 positive and mutant cell lines, including the IL4 [41], EGF [42], NGF [43], CXCR4 [44], IL7 [45], and PDGF [46] pathways, which have all shown roles for p53 in their regulation and signaling. The method that we describe here for identifying prognostic gene sets can provide a more inclusive list of gene sets that provide further insight into the biology of two sample case studies from microarray experiments.
Conclusion
In this paper, we have presented a multiple testing procedure to identify prognostic gene sets from a microarray experiment correlated with common types of binary, continuous and time to event clinical outcomes. We calculate the marginal Pvalues using a permutation method accounting for dependency among the genes within and across each gene set, and account for multiple testing by controlling the FDR. Our simulations show that our proposed method controls the FDR at the desired level. Through extensive simulations and real case studies, we observe that our method performs better than GSEA and GSA, especially when the number of prognostic gene sets is large.
Declarations
Acknowledgements
Partial support for this research was provided by a grant from the National Cancer Institute (CA142538).
Authors’ Affiliations
References
 Ogata H, Goto S, Sato K, Fujibuchi W, Bono H, Kanehisa M: KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Res 1999, 27: 29–34. 10.1093/nar/27.1.29PubMed CentralView ArticlePubMedGoogle Scholar
 Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, IsselTarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G: Gene Ontology: tool for the unification of biology. Nat Genet 2000, 25: 25–29. 10.1038/75556PubMed CentralView ArticlePubMedGoogle Scholar
 Ackermann M, Strimmer K: A general modular framework for gene set enrichment analysis. BMC Bioinformatics 2009, 88: 365–411.Google Scholar
 Mootha VK, Lindgren CM, Eriksson KF, Subramanian A, Sihag S, Lehar J, Puigserver P, Carlsson E, Ridderstrale M, Laurila E, Houstis N, Daly MJ, Patterson N, Mesirov JP, Golub TR, Tamayo P, Spiegelman B, Lander ES, Hirschhorn JN, Altshuler D, Groop LC: PGC1alpharesponsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat Genet 2003, 34: 267–273. 10.1038/ng1180View ArticlePubMedGoogle Scholar
 Goeman JJ, van de Geer SA, de Kort F, van Houwelingen HC: A global test for groups of genes: testing association with a clinical outcome. Bioinformatics 2004, 20: 93–99. 10.1093/bioinformatics/btg382View ArticlePubMedGoogle Scholar
 Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, Mesirov JP: Gene set enrichment analysis: A knowledgebased approach for interpreting genomewide expression profiles. Proc Natl Acad Sci USA 2005, 102: 15545–15550. 10.1073/pnas.0506580102PubMed CentralView ArticlePubMedGoogle Scholar
 Efron B, Tibshirani R: On testing the significance of sets of genes. Annals of Applied Statistics 2007, 1: 107–129. 10.1214/07AOAS101View ArticleGoogle Scholar
 Mansmann U, Meister R: Testing differential gene expression in functional groups. Goeman's global test versus an ANCOVA approach. Methods of Inf Med 2005, 44: 449–453.Google Scholar
 Barry WT, Nobel AB, Wright F: Significance analysis of functional categories in gene expression studies: a structured permutation approach. Bioinformatics 2005, 21: 1943–1949. 10.1093/bioinformatics/bti260View ArticlePubMedGoogle Scholar
 Kong SW, Pu WT, Park PJ: A multivariate approach for integrating genomewide expression data and biological knowledge. Bioinformatics 2006, 22: 2373–2380. 10.1093/bioinformatics/btl401PubMed CentralView ArticlePubMedGoogle Scholar
 Nettleton D, Recknor J, Reecy JM: Identification of differentially expressed gene categories in microarray studies using nonparametric multivariate analysis. Bioinformatics 2008, 24: 192–201. 10.1093/bioinformatics/btm583View ArticlePubMedGoogle Scholar
 Barry WT, Nobel AB, Wright F: A statistical framework for testing functional categories in microarray data. Annals of Applied Statistics 2008, 2: 286–315.View ArticleGoogle Scholar
 Tsai CA, Chen JJ: Multivariate analysis of variance test for gene set analysis. Bioinformatics 2009, 25: 897–903. 10.1093/bioinformatics/btp098View ArticlePubMedGoogle Scholar
 Benjamini Y, Hochber Y: Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society. Series B 1995, 57: 289–300.Google Scholar
 Development Core Team: R: A Language and Environment for Statistical Computing. 2009. ISBN 3900051070Google Scholar
 Wu H, Yang H, Sheppard K, Churchill G, Kerr K, Cui X: maanova: Tools for analyzing Micro Array experiments. R package version 1.20.0 2010.Google Scholar
 Schaefer J, OpgenRhein R, Strimmer K: corpcor: Efficient Estimation of Covariance and (Partial) Correlation. R package version 1.5.7 2010.Google Scholar
 Efron B, Tibshirani R: GSA: Gene set analysis. R package version 1.03 2010.Google Scholar
 Dabney A, Storey JD, Warnes GR: qvalue: Qvalue estimation for false discovery rate control. R package version 1.24.0 2010.Google Scholar
 Gentleman RC, Carey VJ, Bates DM, Bolstad B, Dettling M, Dudoit S, Ellis B, Gautier L, Ge Y, Gentry J, Hornik K, Hothorn T, Huber W, Iacus S, Irizarry R, Leisch F, Li C, Maechler M, Rossini AJ, Sawitzki G, Smith C, Smyth G, Tierney L, Yang JY, Zhang J: Bioconductor: Open software development for computational biology and bioinformatics. Genome Biology 2004, 5: R80. 10.1186/gb2004510r80PubMed CentralView ArticlePubMedGoogle Scholar
 Carlson M, Falcon S, Pages H, Li N: hu6800.db: Affymetrix HuGeneFL Genome Array annotation data (chip hu6800). R package version 2.4.5 2010.Google Scholar
 Cox DR, Hinkley DV: Theoretical Statistics. Chapman and Hall: London; 1974.View ArticleGoogle Scholar
 Warton DI: Penalized normal likelihood and ridge regularization of correlation and covariance matrices. Journal of the American Statistical Association 2009, 103: 340–349.View ArticleGoogle Scholar
 Schafer J, Strimmer K: A shrinkage approach to largescale covariance matrix estimation and implications for functional genomics. Statistical Applications in Genetics and Molecular Biology 2005, 4: 32.View ArticleGoogle Scholar
 Ledoit O, Wolf M: A Wellconditioned estimator for largedimensional covariance matrices. Journal of Multivariate Analysis 2004, 88: 365–411. 10.1016/S0047259X(03)000964View ArticleGoogle Scholar
 Box GEP: Some theorems on quadratic forms applied in the study of analysis of variance problems, I. Effect of inequality of variance in the oneway classification. The Annals of Mathematical Statistics 1995, 25: 290–302.View ArticleGoogle Scholar
 Brunner E: Asymptotic and approximate analysis of repeated measuresd esigns under heteroscedasticity. In mathematical statistics with applications in biometrys. Edited by: Kunert J, Trenkler G. Josef Eul Verlag, Lohmar; 2001.Google Scholar
 Chen SX, Qin YL: A two sample test for high dimensional data with application to geneset testing. The Annals of Statistics 2010, 38: 808–835. 10.1214/09AOS716View ArticleGoogle Scholar
 Cox DR: Regression models and lifetables. Journal of the Royal Statistical Society. Series B 1972, 34: 187–220.Google Scholar
 Lin DY, Wei LJ: The robust inference for the Cox proportinal havards model. Journal of the American Statistical Association 1989, 84: 1074–1078. 10.2307/2290085View ArticleGoogle Scholar
 Storey JD: A direct approach to false discovery rates. Journal of the Royal Statistical Society. Series B 2002, 64: 479–498. 10.1111/14679868.00346View ArticleGoogle Scholar
 Beer DG, Kardia SL, Huang CC, Giordano TJ, Levin AM, Misek DE, Lin L, Chen G, Gharib TG, Thomas DG, Lizyness ML, Kuick R, Hayasaka S, Taylor JM, Iannettoni MD, Orringer MB, Hanash S: Geneexpression profiles predict survival of patients with lung adenocarcinoma. Nat Med 2002, 25: 25–29.Google Scholar
 Barakata TS, Jonkers I, Monkhorst K, Gribnau J: Xchanging information on X inactivation. Exp Cell Res 2010, 316: 679–687. 10.1016/j.yexcr.2010.01.015View ArticleGoogle Scholar
 Prothero KE, Stahl JM, Carrel L: Dosage compensation and gene expression on the mammalian X chromosome: one plus one does not always equal two. Chromosome Res 2009, 17: 637–648. 10.1007/s1057700990639View ArticlePubMedGoogle Scholar
 Zhang W, Huang RS, Duan S, Dolan ME: Gene set enrichment analyses revealed differences in gene expression patterns between males and females. In Silico Biol 2009, 9: 55–63.PubMed CentralPubMedGoogle Scholar
 Parini P, Jiang ZY, Einarsson C, Eggertsen G, Zhang SD, Rudel LL, Han TQ, Eriksson M: ACAT2 and human hepatic cholesterol metabolism: identification of important genderrelated differences in normolipidemic, nonobese Chinese patients. Atherosclerosis 2009, 207: 266–271. 10.1016/j.atherosclerosis.2009.04.010PubMed CentralView ArticlePubMedGoogle Scholar
 Bogani D, Siggers P, Brixey R, Warr N, Beddow S, Edwards J, Williams D, Wilhelm D, Koopman P, Flavell RA, Chi H, Ostrer H, Wells S, Cheeseman M, Greenfield A: Loss of mitogenactivated protein kinase kinase kinase 4 (MAP3K4) reveals a requirement for MAPK signalling in mouse sex determination. PLoS Biol 2009, 7: e1000196. 10.1371/journal.pbio.1000196PubMed CentralView ArticlePubMedGoogle Scholar
 Yamasaki K, Kurimura M, Kasai T, Sagara M, Kodama T, Inoue K: Determination of physiological plasma pentraxin 3 (PTX3) levels in healthy populations. Clin Chem Lab Med 2009, 47: 471–477. 10.1515/CCLM.2009.110View ArticlePubMedGoogle Scholar
 Khymenets O, Covas MI, Farre M, Langohr K, Fito M, de la Torre R: Role of sex and time of blood sampling in SOD1 and SOD2 expression variability. Clin Biochem 2008, 41: 1348–1354. 10.1016/j.clinbiochem.2008.08.064View ArticlePubMedGoogle Scholar
 Tomasini R, Mark TW, Melino G: The impact of p53 and p73 on aneuploidy and cancer. Trends Cell Biol 2008, 18: 244–252. 10.1016/j.tcb.2008.03.003View ArticlePubMedGoogle Scholar
 Pesch J, Brehm U, Staib C, Grummt F: Repression of interleukin2 and interleukin 4 promoters by tumor suppressor protein p53. J Interferon Cytokine Res 1996, 16: 595–600. 10.1089/jir.1996.16.595View ArticlePubMedGoogle Scholar
 Sheikh MS, Carrier F, Johnson AC, Ogdon SE, Fornace AJ Jr: Identification of an additional p53responsive site in the human epidermal growth factor receptor gene promotor. Oncogene 1997, 15: 1095–1101. 10.1038/sj.onc.1201264View ArticlePubMedGoogle Scholar
 Brynczka C, Labhart P, Merrick BA: NGFmediated transcriptional targets of p53 in PC12 neuronal differentiation. BMC Genomics 2007, 8: 139. 10.1186/147121648139PubMed CentralView ArticlePubMedGoogle Scholar
 Mehta SA, Christopherson KW, BhatNakshatri P, Goulet RJ Jr, Broxmeyer HE, Kopelovich L, Nakshatri H: Negative regulation of chemokine receptor CXCR4 by tumor suppressor p53 in breast cancer cells: implications of p53 mutation or isoform expression on breast cancer cell invasion. Oncogene 2007, 26: 3329–3337. 10.1038/sj.onc.1210120View ArticlePubMedGoogle Scholar
 Costello PS, Cleverley SC, Galandrini R, Henning SW, Cantrell DA: The GTPase rho controls a p53dependent survival checkpoint during thymopoiesis. J Exp Med 2000, 192: 77–85. 10.1084/jem.192.1.77PubMed CentralView ArticlePubMedGoogle Scholar
 Yang W, Wetterskog D, Matsumoto Y, Funa K: Kinetics of repression by modified p53 on the PDGF betareceptor promoter. Int J Cancer 2008, 123: 2020–2030. 10.1002/ijc.23735View ArticlePubMedGoogle Scholar
Copyright
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.