• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of bioinfoLink to Publisher's site
Bioinformatics. Oct 1, 2009; 25(19): 2588–2594.
Published online Aug 11, 2009. doi:  10.1093/bioinformatics/btp439
PMCID: PMC2752611

Testing the hypothesis of tissue selectivity: the intersection–union test and a Bayesian approach

Abstract

Motivation: Finding genes that are preferentially expressed in a particular tissue or condition is a problem that cannot be solved by standard statistical testing procedures. A relatively unknown procedure that can be used is the intersection–union test (IUT). However, two disadvantages of the IUT are that it is conservative and it conveys only the information of the least differing target tissue–other tissue pair.

Results: We propose a Bayesian procedure that quantifies how much evidence there is in the overall expression profile for selective over-expression. In a small simulation study, it is shown that the proposed method outperforms the IUT when it comes to finding selectively expressed genes. An application to publicly available data consisting of 22 tissues shows that the Bayesian method indeed selects genes with functions that reflect the specific tissue functions. The proposed method can also be used to find genes that are underexpressed in a particular tissue.

Availability: Both MATLAB and R code that implement the IUT and the Bayesian procedure in an efficient way, can be downloaded at http://ppw.kuleuven.be/okp/software/BayesianIUT/.

Contact: katrijn.vandeun/at/psy.kuleuven.be

1 INTRODUCTION

Gene expression is often profiled for several tissues to get insights into gene function and regulation (Dezso et al., 2008; Liu et al., 2008; Su et al., 2004). In this respect, finding genes that are selectively expressed in a particular tissue is of importance to unravel the biological processes taking place in the particular tissue and to identify candidate biomarkers (Klee, 2008; Liang et al., 2006; Su et al., 2004). It is important to make a distinction between three types of tissue-selectivity [see Klee (2008) for a review]: (i) the gene is only expressed in a particular tissue; usually this type of expression is called ‘tissue-specific’ (Skrabanek and Campagne, 2001); (ii) the gene is expressed at approximately the same level in all tissues except one [called categorical tissue specificity by Schug et al. (2005)]; or (iii) the gene is over- or underexpressed in a particular tissue compared with all the other tissues (Kadota et al., 2003, 2006). The latter broadest type of preferential expression in a particular tissue is often called tissue-selective (Greller and Tobin, 1999; Liang et al., 2006). Note that tissue-specific as defined by Skrabanek and Campagne (2001) is a special case of categorical tissue specificity and that categorical tissue specificity is a special case of tissue-selectivity. In accordance with these different definitions of tissue-selectivity, different methods to find genes with the particular expression profile have been proposed. The remainder of the article focuses on the broadest class of tissue-selectivity, namely relative over- or underexpression in a particular tissue compared with all the other tissues. All methods are explained and illustrated for selective over-expression, but the generalization to selective underexpression is straightforward.

To find tissue-selective genes, we found two approaches in the literature that rely on a firm statistical framework, hereby reducing arbitrary choices to a minimum. A first approach is unsupervised and relies on outlier detection to scan expression profiles for outlying values (Kadota et al., 2003, 2006). The resulting tissue-selective profiles can be selectively expressed in more than one tissue and can be both up- or downregulated in these tissues. A drawback of the method is that it cannot account for biological and technical variation because either all replicate values are included in the analysis (with the likely outcome that only some replicate values will be detected as outlying), or a single representative measure (e.g. average over the replicates) has to be used in the analysis. The second approach is supervised and constructed for the case of replicate arrays for each tissue (Liang et al., 2006). It relies on hypothesis testing procedures to test whether a gene is selectively overexpressed in a particular tissue. t-tests are used to measure how significant the difference in expression of each of the target tissue–other tissue pairs is and a gene is declared tissue-selective when each of the differences is significant. The problem of multiple testing is accounted for by using the Tukey–Kramer multiple comparison procedure. Although the obtained tissue-selective genes are claimed to have higher expression in the target tissue than in each of the other tissues, the statistical procedure used is tailored to find those genes that are significantly higher expressed in the target compared with at least one of the other tissues. To understand why this is the case, observe that Tukey–Kramer and other common multiple comparison procedures like Bonferroni and Dunnett (Dunnett, 1955) control the chance of wrongfully rejecting the null hypothesis of no significantly different target–other tissue pair against the alternative of at least one significantly different pair. What is needed, is a test that controls the chance of wrongfully rejecting the null hypothesis against the alternative of all significantly different target–other tissue pairs.

A test that would have the desired alternative hypothesis of significantly higher expression in the target tissue compared with each of the other tissues, was proposed by Berger (1982) and is known as the intersection–union test (IUT; Berger and Hsu, 1996). However, the IUT test has two disadvantages. First, it is conservative implying that many tissue-selective genes would be missed; and second, the obtained results are not very informative as the test only indicates whether a gene is tissue-selective or not without any distinction in the degree of tissue selectivity (see also Allison et al., 2006). As an alternative, we would like to propose a Bayesian procedure.

In this article, we first briefly describe the IUT and we introduce a Bayesian alternative. Subsequently, both procedures are compared in a simulation study and the Bayesian procedure is used to find the tissue-selective genes for a panel of 22 tissues.

2 METHODS

2.1 Intersection–union test

Assume some specific gene and a specific target tissue t. Let us further denote by H0j the partial null hypothesis that the gene under study is expressed equally or higher in tissue j than in the target tissue. Furthermore, we denote by H1j the partial alternative that that the gene is expressed higher in the tissue target t. Then the gene is selectively upregulated in target tissue t, if the compound null hypothesis that H0j holds for at least one tissue (jt), is rejected against the compound alternative hypothesis that for all tissues (jt) H1j holds. Using formal notation, the set of compound hypotheses is composed by An external file that holds a picture, illustration, etc.
Object name is btp439i1.jpg and An external file that holds a picture, illustration, etc.
Object name is btp439i2.jpg. Note that this is different from common multiple comparison procedures where the underlying set of compound hypotheses is that all partial null hypotheses hold (i.e. An external file that holds a picture, illustration, etc.
Object name is btp439i3.jpg) against the alternative that at least one partial null hypothesis can be rejected (i.e. (An external file that holds a picture, illustration, etc.
Object name is btp439i4.jpg)). Berger (1982) introduced a procedure to test the composite null hypothesis An external file that holds a picture, illustration, etc.
Object name is btp439i5.jpg against the composite alternative An external file that holds a picture, illustration, etc.
Object name is btp439i6.jpg (see also Berger and Hsu, 1996). A result is declared significant by this test at level α, if it holds that each partial null hypothesis H0j can be rejected at level α. As proven by Berger, the significance level of his test is less than or equal to the significance level used for each of its implied partial tests.

The IUT can be applied to the problem of finding tissue-selective upregulated genes as follows. For a particular gene, test each target tissue–other tissue pair at the desired significance level (e.g. 0.05) using a suitable test-statistic like the t-test. Only when all pairs yield a significant result, the gene is declared selectively overexpressed. Note that to account for the problem of testing multiple genes, the significance level used can be adapted using Bonferonni's or Sidak's correction. An adaptation to finding underexpressed genes is straightforward by testing the partial null hypothesis that the gene under study is expressed equally or lower in tissue j than in the target tissue against the partial alternative that the gene is expressed lower in the target tissue. An efficient implementation of this approach in MATLAB or R can be found online (http://ppw.kuleuven.be/okp/software/BayesianIUT/).

Often the significance level of the IUT is (much) less than α such that the procedure is conservative (a gene will not be easily declared to be selectively upregulated in the tissue). This yields a very low false discovery rate, however, at the cost of many false negatives. Deng et al. (2008) proposed an adjusted IUT for the special case of two independent tests that is less conservative. Note, however, that it is not suitable to find tissue-selective genes (the common target makes that the tests are dependent and usually interest is in comparing the target with more than two other tissues). Another disadvantage of the IUT is that it indicates whether a gene is selectively overexpressed or not, but not to which degree. No distinction is made, for example, at the 0.05 level of significance between a tissue-selective upregulated gene that has P-value of 0.03 for each of the target–other tissue pairs and a gene that has P-value of 0.0001. The IUT can be made somewhat more informative by reporting the largest P-value of the partial tests (Tuke et al., 2009), thus focusing only on the least differing target tissue–other tissue pair.

2.2 Bayesian evaluation of the constrained hypothesis

In this section, we describe a Bayesian alternative for the IUT to evaluate the hypothesis of tissue-selective overexpression. Note that the procedure will be explained for one gene. The data consist of i = 1,…, N expression levels yi for j = 1,…, J tissues. The total sample size N = ∑j Nj with Nj denoting the number of replications for tissue j which means that i is nested in j. So, for example, for j = 1, i = 1,…, N1, for j = 2, i = N1 + 1,…, N1 + N2. The model for the expression level is

equation image
(1)

where μj denotes the population mean of tissue j, dji is 1 if the expression was obtained for tissue j and 0 otherwise and ϵi ~ [mathematical script N](0, σ2). The density of the data for this ANOVA model is

equation image
(2)

where y=[y1,…, yN] and dj = [dj1,…, djN].

The goal is to determine the support in the data for two hypotheses:

equation image
(3)

which states that μ1 is larger than each of the means in the set {μ2,…, μJ} (the gene is tissue-selective), and

equation image
(4)

Note that H2 corresponds to H0 of the IUT. Support in the data will be quantified using the Bayes factor (Kass and Raftery, 1995). Using Chib's approach (Chib, 1995) the Bayes factor of H1 versus H2 can be written as:

equation image
(5)

where θ=[μ1,…, μJ], h(·) denotes the prior distribution of the parameters for the hypothesis indicated and g(·) the posterior distribution. The dependence of f(·) and g(·) on d1,…, dJ is left implicit. Note that BF12=6 implies that the support in the data for H1 is six times as large as the support in the data for H2.

It is convenient to write

equation image
(6)

where u refers to Hu : μ1,…, μJ, that is, an unconstrained model. Once a prior distribution has been specified for Hu the prior distribution of Hm for m = 1, 2 is obtained via

equation image
(7)

where Iθ[set membership]m = 1 if θ is in agreement with the constraints of model m and 0 otherwise, and cm is the proportion of Hu in agreement with Hm.

The prior distribution of Hu is

equation image
(8)

that is, the prior distribution of each μj is the same. With this specification, and independent of the choice of μ0, τ0, ν0 and σ20, c1 = 1/J and c2 = (J − 1)/J, which makes sense because there are J equivalent models in which one of the means is larger than the other means. Using (7) and (8), BF1u can, for any value of θ in agreement with H1, be written as:

equation image
(9)

where f1 is the proportion of the unconstrained posterior distribution in agreement with the constraints of H1 because:

equation image
(10)

Using a similar derivation for BF2u, it is obtained that

equation image
(11)

See Klugkist and Hoijtink (2007) for a more elaborate discussion of the derivation of the Bayes factor for inequality constrained hypotheses in the context of ANOVA. As will be illustrated in Section 3.1, if the number of means smaller than the target mean increases, the Bayes factor increases in favor of the hypothesis of selective overexpression and if the number of means larger than the target mean increases, the Bayes factor increases in favor of the complement of the hypothesis of selective overexpression. Furthermore, note the following property of the Bayes factor. Consider the situation where all means have the same value. Then the expected value of f1 = 1/J and f2 = (J − 1)/J, that is, BF12 = 1. Stated otherwise, if all means are equal the Bayes factor is neutral with respect to the hypothesis of interest and its complement. According to the Bayes factor one or more means being equal to the mean of the target tissue is neither evidence in favor nor against the hypothesis of selective overexpression.

The only question remaining is the estimation of f1 because f2=1 −f1. Using a very large (but finite) number for τ0, any number for μ0, ν = −2 and σ20 = 0, that is very uninformative priors, the following algorithm renders a sample from the unconstrained posterior distribution of μ1,…, μJ, σ2 and an estimate of f1:

  • Step 1: assign initial values: An external file that holds a picture, illustration, etc.
Object name is btp439i7.jpg for j=1,…, J and An external file that holds a picture, illustration, etc.
Object name is btp439i8.jpg, where An external file that holds a picture, illustration, etc.
Object name is btp439i9.jpg denotes the sample average for tissue j.
  • Step 2: for j=1,…, J sample μj from gj [mid ] σ2, y) which is a normal distribution with mean An external file that holds a picture, illustration, etc.
Object name is btp439i10.jpg and variance σ2/Nj.
  • Step 3: verify whether or not the current values of μ1,…, μJ are in agreement with the constraints of H1.
  • Step 4: sample σ2 from g2 [mid ] μ1,…, μJ, y) which is a scaled inverse chi-square distribution with degrees of freedom N−2 and scale parameter 1/(N−2)∑i(yi−μ1d1i−···−μJdJi)2.
  • Iterate Steps 2 through 4. The proportion of vectors μ1,…, μJ sampled in Step 3 in agreement with the constraints of H1 is an estimate of f1.

Note that this algorithm has two favorable properties. First, our approach is objective in the following sense: (i) due to the use of vague prior distributions, the posterior is proportional to the likelihood, that is, f1 and f2 are completely determined by the data, and (ii) since the prior is the same for each mean, c1 and c2 do not depend on the prior. Second, no burn-in is needed because convergence is almost instantaneously as we sample from an inverse normal chi-square distribution and because we initialize the algorithm with the sample average and sample variance. In applications, we used 5000 iterations in the Gibbs sampler. Testing for selective underexpression can be done by adapting the constraints in H1; on the algorithmic level this only influences Step 3. An efficient implementation of this algorithm in MATLAB or R can be found online (http://ppw.kuleuven.be/okp/software/BayesianIUT/).

3 RESULTS

We compared the performance of the IUT and Bayesian procedure in finding tissue-selective overexpressed genes using simulated and real data. Real data were obtained after robust multichip analysis (RMA) preprocessing and a log2 transformation. Because the simulation can be used as a reference for the empirical data, several parameters in the simulation were chosen the same as for these data. These are the number of tissues (22), the number of replications per tissue (three to five), the total number of expression values for a gene (70), the SD per tissue calculated over the replications (s=0.04, corresponding to the median SD in the data) and the overall mean expression level (μ = 7.3).

3.1 Simulation

Here, we compare the performance of the two testing procedures in a controlled setting using simulated data. Two influencing factors are of interest, namely the amount of support in the data and the degree to which the expression differs between the target and other tissues (the so-called effect size; Cohen, 1969). First, for the amount of support four levels will be considered: (i) complete support by all 21 tissues to which the target is compared; (ii) one tissue that is neutral, that is, has the same mean as the target tissue, and 20 supporting; (iii) one tissue not supporting and 20 supporting; and (iv) none of the 21 tissues supporting. Second, the effect size δ is manipulated at two levels: one with a considerable overlap of the sampling distributions, δ = 0.5s = 0.02, and one with a small overlap δ = 2s = 0.08. The data are generated from a normal distribution with SD s = 0.04 and mean equal to 7.3 for the target, to 7.3 − δ for a tissue supporting the hypothesis and to 7.3 + δ for a tissue not supporting the hypothesis. Note that the effect sizes used may seem too small for what can be expected from expression data. However, using effect sizes larger than approximately three leads to an almost complete separation of the sampling distributions such that performance of both the IUT and Bayesian approach would be almost perfect.

The results for the simulation are summarized in Table 1 that reports the proportion of genes for which the data support that they are selectively overexpressed in the target tissue. The panels correspond to different decision rules for tissue-selective overexpression: in the first panel, this is a significant result for the IUT; in the second panel, this is a Bayes factor larger than one (more support for H1 than for H2); and in the third panel this is for a Bayes factor larger than 32 (support for H1 is 32 times larger than the support for H2). The rows correspond to the four different amounts of support, while the two columns of each panel correspond to the two levels of effect size. Histograms of the log transformed Bayes factor1 are shown in Figure 1. These were used to determine the cutoff of 32 (corresponding to a log2 value equal to 5).

Table 1.
Proportion of tissue-selective overexpressed genes
Fig. 1.
Histograms of the log2 transformed Bayes factors. (A) Complete support of H1. (B) One tissue neutral with respect to H1. (C) One tissue not supporting H1. (D) None of the tissues supporting H1. Panels at the left are obtained with a considerable overlap ...

Clearly, the IUT is conservative. Even in the most favorable case (complete support, large effect size), only 26% of the genes are declared to be tissue-selective,2 corresponding to a false negative rate of 74%. Compare this with the corresponding results for the Bayesian procedure (middle panel of Table 1): the majority of the Bayes factors indicate that H1 is supported, even when the effect size is small. As can be seen in the third row of Table 1, this more liberal character of the Bayesian approach leads to many false positive results, but can be solved by requiring that the Bayes factor should be >32. In this case, as shown in the right panel of Table 1, the false positive rate is close to zero yet the false negative rate is much smaller than for the IUT. The second row of Table 1 presents the situation where one tissue has the same expression level as the target, and the other expression levels are smaller than the expression level of the target tissue. As elaborated in the previous section for the Bayesian approach, expression levels equal to the expression level of the target tissue are neither evidence in favor nor against the hypothesis of interest. This is reflected by an increased rate of genes detected as selectively overexpressed for both the IUT and Bayes factor. Taking 32 as a demarcation value for the Bayes factor yields 30% of the genes detected as selectively overexpressed. In case that the pair of tissues cannot be considered as a functionally equivalent group, these are false positives and avoiding them can be solved by taking a much higher Bayes factor (e.g. log2 = 10, see Fig. 1B) as a demarcation value. Then, also the Bayesian procedure becomes conservative. Note that adding more tissues with an expression level equal to the target tissue, will quickly reduce the proportion supported to almost zero (e.g. with two tissues having the same expression as the target, the 0.30 drops to 0.16). As illustrated in Table 1, this holds also for a lower effect size (δ = 0.5s) for the tissues that support the hypothesis H1. The fourth row shows that the performance of both the IUT and the Bayes factor is rather good if there is no support at all for the hypothesis of interest.

To illustrate that the Bayesian procedure is more informative than reporting the largest P-value of the partial tests involved in the IUT, we simulated data for which one tissue has the same mean as the target tissue (7.3). The means of the remaining tissues steadily varied from 7.3 to 7.3 − 3s = 7.3 − 0.12 (corresponding to an effect size that varies from 0 to 3), this is increasingly supporting the hypothesis of selective overexpression in the target tissue. For each effect size, 50 replicates (genes) were generated. A scatterplot for the IUT is depicted in Figure 2A, where the largest P-value of the partial tests is plotted against the effect size for the IUT; a scatterplot for the Bayesian procedure is depicted in Figure 2B where the log of the Bayes factor is plotted against this same effect size. Clearly, the Bayesian procedure detects that the support of the 20 tissues increases, while this information is not captured by the IUT.

Fig. 2.
Scatter plots for the largest P-value of the IUT (A) and of the log2 Bayes factors (B) in function of the effect size.

Advantages of the IUT over the Bayesian approach are that it is less computationally intensive and has a clear cutoff. First, running the simulation discussed above on an Intel Core Duo took less than a second for the IUT and less than 10 min for the Bayesian approach (with 5000 iterations): this is for all eight conditions and with 5000 replications per condition (in practice, this corresponds to eight analyses of 5000 genes). Note that although the Bayesian procedure is much slower, the time required for the analysis of a large dataset using an ordinary desktop is still very reasonable. Second, the conventional use of 0.01 or 0.05 as a cutoff for significance gives a clear rule. For the Bayes factor, there is not such a clear rule: Choosing a cutoff that yields a good balance between the number of true and false positives (this is not too conservative nor too liberal) depends on the number of tissues, their number of replicates and their effect sizes. Therefore, a simulation study should be performed prior to the analysis of the intended data; as an aid for potential users, we provided the script file of our simulation.

3.2 Tissue-selective genes

We used a publicly available microarray dataset (http://www.ncbi.nlm.nih.gov/geo/, accession number GSE9954; see also Thorrez et al., 2008) that we generated via Affymetrix mRNA expression analysis using 430 2.0 arrays. This database consists of 22 different murine tissues, with 3–5 replicates for each tissue. Note that in the Bayesian procedure, the replications are supposed to follow a normal distribution; therefore, we took the log2 of the RMA preprocessed expression data.

For each tissue, the Bayesian procedure was used to find genes preferentially expressed in a particular tissue. We decided that a gene is tissue-selective upregulated when the Bayes factor was >32 on all probesets matching the gene. Genes that are tissue-selective are expected to be associated with the cellular processes which are the characteristic for the tissue. To assess whether this was the case for the gene sets identified by the Bayesian procedure, we tested functional overrepresentation of these genes using Ingenuity Pathway Analysis 7.1. The five most significant functions and diseases per tissue are listed in Table 2. It is apparent that most of these clearly reflect tissue-specific functions. This indicates that the underlying gene sets truly capture the tissue selectivity. For seminal vesicle only two functions reached significance and for salivary gland no significant function was found. Probably, this is due to the fact that these two tissues are studied by few researchers, leading to few publications on which Ingenuity Pathway Analysis can base its results.

Table 2.
For each tissue, the genes identified by the Bayesian procedure to be tissue-selective were analyzed for functional overrepresentation

It is important to realize that the obtained results depend strongly on the panel of tissues considered. Including more tissues will lead to a smaller number of genes denoted to be tissue-selective by the Bayesian procedure, but the biological specificity for these genes will be higher. For example, the panel used here contains three contractile tissues: gastrocnemius, diaphragm and heart; especially gastrocnemius and diaphragm have rather similar functions. Therefore, the number of tissue-selective genes is small (e.g. for gastrocnemius only 132 genes were found; see Table 2) and their functions are very specific (e.g. quantity of skeletal muscle associated to gastrocnemius and rib formation to diaphragm; see Table 2).

4 DISCUSSION

Finding tissue-selective genes is a recurring biological theme. As shown, the use of procedures that correct for multiple testing associated to the comparison of all target tissue–other tissue pairs is erroneous. A correct statistical procedure is the IUT. However, it is conservative and limited in information. As an alternative, we proposed a Bayesian procedure. In a simulation study, it was shown that this procedure is in most situations less conservative, while still keeping the number of false positives acceptable. Also, it is more informative than the IUT because it expresses how strongly the complete expression profile supports the hypothesis of tissue selectivity. Note that although we discussed the case of selective overexpression, both methods can also be used to find selective underexpressed genes.

The results of both the IUT and Bayesian procedure are highly dependent on the panel of tissues considered. These methods denote a gene as preferentially expressed in a particular tissue when expression in the tissue is higher than in each of the other tissues considered in the panel. Small panels can be expected to lead to large lists of tissue-selective genes containing many genes that are even more selectively overexpressed in a tissue not included in the panel. In an application of the Bayesian procedure to a panel of 22 tissues, we illustrated this by including two tissues with rather similar functionality, namely gastrocnemius and diaphragm. To avoid biologically flawed results, the proposed Bayesian method as well as the IUT should be applied with careful consideration of the tissues to include in the panel.

The use of the Bayesian procedure and the IUT is not limited to finding tissue-selective genes in normal tissues. Any problem involving the comparison of a reference group to each of the other groups (more than two) can be envisaged. Interesting applications are the comparison of normal tissue (persons) to several types of diseased tissue (persons). For example, Nishimura et al. (2007) aimed at genes susceptible of autism by comparing normal persons to each of two groups of autistic persons.

An advantage of the Bayesian approach is that it is flexible in the kind of hypotheses that can be tested with it. Not only can it be easily adapted to test the hypothesis H1 of selective expression in a few (and not a single) tissue against the complement H2: not H1, but also against more specific alternative hypotheses H2. For example, it may be of interest to know whether the data support that a particular gene is upregulated in a very specific tissue belonging to a group of related tissues (e.g. three neuronal tissues) rather than that it is upregulated overall in this group of related tissues . The procedure that was proposed here is general and can thus be easily adapted to test such hypotheses. On the other hand, the IUT always tests against H2: not H1. A more challenging adaptation of the Bayesian procedure would be the possibility to include equalities in the hypothesis (e.g. to find categorical tissue-specific genes) in a way that the procedure is efficient enough to deal with thousands of genes [see Klugkist and Hoijtink (2007) for a discussion of Bayes factors for equality and inequality constrained hypotheses]. Inclusion of such equality constraints for the IUT are discussed by Tuke et al. (2009) (an equivalence testing approach is used).

ACKNOWLEDGEMENTS

We wish to thank two anonymous reviewers for helpful comments.

Funding: the Research Fund of Katholieke Universiteit Leuven (SymBioSys: CoE EF/05/007 and GOA/2005/04).

Conflict of Interest: none declared.

Footnotes

1Bayes factors below 0.0001 were set equal to 0.0001, while Bayes factors over 10 000 were set equal to 10 000.

2This number drops to 0 when accounting for the multiple testing associated to the testing of 5000 genes (results not shown).

REFERENCES

  • Allison DB, et al. Microarray data analysis: from disarray to consolidation and consensus. Nat. Rev. Genet. 2006;7:55–65. [PubMed]
  • Berger RL, Hsu JC. Bioequivalence trials, intersecion-union tests, and equivalence confidence sets. Stat. Sci. 1996;11:283–319.
  • Berger RL. Multiparameter hypothesis testing and acceptance sampling. Technometrics. 1982;24:295–300.
  • Chib S. Marginal likelihood from Gibbs output. J. Am. Stat. Assoc. 1995;90:1313–1321.
  • Cohen J. Statistical Power Analysis for the Behavioral Sciences. New York: Academic Press; 1969.
  • Deng X, et al. Improving the power for detecting overlapping genes from multiple DNA microarray-derived gene lists. BMC Bioinformatics. 2008;9(Suppl. 6):S14. [PMC free article] [PubMed]
  • Dezso Z, et al. A comprehensive functional analysis of tissue specificity of human gene expression. BMC Biol. 2008;6:49. [PMC free article] [PubMed]
  • Dunnett CW. A multiple comparison procedure for comparing several treatments with a control. J. Am. Stat. Assoc. 1955;50:1096–1121.
  • Greller LD, Tobin FL. Detecting selective expression of genes and proteins. Genome Res. 1999;9:282–296. [PMC free article] [PubMed]
  • Kadota K, et al. Detection of genes with tissue-specific expression patterns using Akaike's information criterion procedure. Physiol. Genomics. 2003;12:251–259. [PubMed]
  • Kadota K, et al. ROKU: a novel method for identification of tissue-specific genes. BMC Bioinformatics. 2006;7:294. [PMC free article] [PubMed]
  • Kass R, Raftery A. Bayes factors. J. Am. Stat. Assoc. 1995;90:773–795.
  • Klee EW. Data mining for biomarker development: a review of tissue specificity analysis. Clin. Lab. Med. 2008;28:127–143. [PubMed]
  • Klugkist I, Hoijtink H. The Bayes factor for inequality and about equality constrained models. Comput. Stat. Data Anal. 2007;51:6367–6379.
  • Liang S, et al. Detecting and profiling tissue-selective genes. Physiol. Genomics. 2006;26:158–162. [PubMed]
  • Liu X, et al. TIGER: a database for tissue-specific gene expression and regulation. BMC Bioinformatics. 2008;9:271. [PMC free article] [PubMed]
  • Nishimura Y, et al. Genome-wide expression profiling of lymphoblastoid cell lines distinguishes different forms of autism and reveals shared pathways. Hum. Mol. Genet. 2007;16:1682–1698. [PubMed]
  • Schug J, et al. Promoter features related to tissue specificity as measured by Shannon entropy. Genome Biol. 2005;6:R33. [PMC free article] [PubMed]
  • Skrabanek L, Campagne F. Tissueinfo: high-throughput identification of tissue expression profiles and specificity. Nucleic Acids Res. 2001;29:E102. [PMC free article] [PubMed]
  • Su AI, et al. A gene atlas of the mouse and human protein-encoding transcriptomes. Proc. Natl Acad. Sci. USA. 2004;101:6062–6067. [PMC free article] [PubMed]
  • Thorrez L, et al. Using ribosomal protein genes as reference: a tale of caution. PLoS ONE. 2008;3:e1854. [PMC free article] [PubMed]
  • Tuke J, et al. Gene profiling for determining pluripotent genes in a time course microarray experiment. Biostatistics. 2009;10:80–93. [PubMed]

Articles from Bioinformatics are provided here courtesy of Oxford University Press

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

  • PubMed
    PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...