A comparative review of estimates of the proportion unchanged genes and the false discovery rate
- Per Broberg^{1}Email author
https://doi.org/10.1186/1471-2105-6-199
© Broberg; licensee BioMed Central Ltd. 2005
Received: 22 December 2004
Accepted: 08 August 2005
Published: 08 August 2005
Abstract
Background
In the analysis of microarray data one generally produces a vector of p-values that for each gene give the likelihood of obtaining equally strong evidence of change by pure chance. The distribution of these p-values is a mixture of two components corresponding to the changed genes and the unchanged ones. The focus of this article is how to estimate the proportion unchanged and the false discovery rate (FDR) and how to make inferences based on these concepts. Six published methods for estimating the proportion unchanged genes are reviewed, two alternatives are presented, and all are tested on both simulated and real data. All estimates but one make do without any parametric assumptions concerning the distributions of the p-values. Furthermore, the estimation and use of the FDR and the closely related q-value is illustrated with examples. Five published estimates of the FDR and one new are presented and tested. Implementations in R code are available.
Results
A simulation model based on the distribution of real microarray data plus two real data sets were used to assess the methods. The proposed alternative methods for estimating the proportion unchanged fared very well, and gave evidence of low bias and very low variance. Different methods perform well depending upon whether there are few or many regulated genes. Furthermore, the methods for estimating FDR showed a varying performance, and were sometimes misleading. The new method had a very low error.
Conclusion
The concept of the q-value or false discovery rate is useful in practical research, despite some theoretical and practical shortcomings. However, it seems possible to challenge the performance of the published methods, and there is likely scope for further developing the estimates of the FDR. The new methods provide the scientist with more options to choose a suitable method for any particular experiment. The article advocates the use of the conjoint information regarding false positive and negative rates as well as the proportion unchanged when identifying changed genes.
Background
The microarray technology permits the simultaneous measurement of the transcription of thousands of genes. The analysis of such data has however turned out to be quite a challenge. In drug discovery, one would like to know what genes are involved in certain pathological processes, or what genes are affected by the intervention of a particular compound. A more basic question is 'How many genes are affected or changed?' It turns out that the answer to this basic question has a bearing on the other questions.
The proportion unchanged
In the two-component model for the distribution of the test statistic the mixing parameter p_{0}, which represents the proportion unchanged genes, is not estimable without strong distributional assumptions, see [1]. Assuming this model the probability density function (pdf) f^{ t }of a test statistic t may be written as the weighted sum of the null distribution pdf and the alternative distribution pdf
If, on the other hand, we know the value of p_{0} we can estimate e.g. through a bootstrap procedure as described in [1], and thus obtain also .
- 1)
Knowing the proportion changed genes in a microarray experiment is of interest in its own right. It gives an important global measure of the extent of the changes studied.
- 2)
The next example concerns FDR. Suppose we reject null hypothesis j, and call gene j significantly regulated, when the corresponding p-value p_{ j }falls below some cutpoint α. The question that motivates the FDR concept, which originates from [2], is: "What proportion of false positives is expected among the selected genes?" A goal would then be to quantify this proportion, and one possible estimate is
- 3)
Knowing p_{0}, we may calculate the posterior probability of a gene being a Differentially Expressed Gene (a DEG) as
- 4)
Knowing p_{0}, it is also possible to estimate the number of false positives and false negatives at a given cutpoint α as a proportion of the total number of genes. Call these proportions the false-positive and false-negative rates, and denote them by FP(α) and FN(α), respectively. In the samroc methodology [5] one calculates estimates of these quantities as
and
One may choose a p-value threshold α_{ min }, which minimises the amount of errors FP(α) + FN(α). Alternatively, one may want to fine-tune the test statistic such that it will minimise the errors at a given threshold. Or, one may try to do both, as suggested in [5], see also [6].
Earlier research providing estimates of p_{0} include [1, 3, 7–13]. Articles that compare methods for estimating p_{0} and FDR include [12, 14]. I will focus on the FDR as the main use of p_{0}.
In this article, the formulation of the theory is in terms of p-values rather than in terms of test statistics. Two basic assumptions are made concerning their distribution. First, it is assumed that test statistics corresponding to true null hypotheses will generate p-values that follow a uniform distribution on the unit interval, e.g. [15]. Thus, under the null distribution, the probability that a p-value falls below some cutpoint α equals α. Second, p-values are, unless stated otherwise, assumed to be independent. Empirical investigations will assess the effects of deviations from the second assumption.
The use of p-values means lumping up- and downregulated genes together. However, one may look separately at the two tails of the distribution of the test statistic to assess differential expression corresponding to up- and downregulation.
This article will not concern how p-values are calculated, but rather how they are used to calculate estimates of p_{0} and FDR, and draw conclusions based on this evidence. It is assumed that p-values capture the essence of the research problem. Neither does the article treat the choice of an optimal test statistic. For illustration the t-test will be used repeatedly without regard to whether there are better methods or not. By the t-test we mean the unequal variance t-test: for sample means mean_{1} and mean_{2}, sample variances , , and sample sizes n_{1} and n_{2}. We apply the t-test to simulated normally distributed data and a permutation t-test to real data, where normality may be uncertain. All calculations were performed in R. The methods presented are available within packages for the free statistical software R [16, 17] and take a vector of p-values as input and output an estimate of p_{0} and of FDR. Emphasis lies on methods available within R packages downloadable from CRAN [18] or Bioconductor [19]. Inevitably any review will exclude interesting work, but time and space limitations will not permit an all comprehensive review. The new and highly interesting concept of a local false discovery rate (LFDR) [1] only receives a cursory treatment.
This article builds on and finds motivation from the experience of the analysis of microarrays, which typically assay the expression of 10,000 or more genes. However, the methods presented apply equally well to other high dimensional technologies, such as fMRI or Mass Spectrometry.
False discovery rate
In the analysis of microarray experiments, the traditional multiple test procedures are often considered too stringent, e.g. [20] and [3]. In the last decade alternatives based on the concept of an FDR have emerged. For more details consult e.g. [2, 3, 9, 11, 21, 22]. There are different definitions proposed, but loosely speaking one would want to measure the proportion of false positive genes among those selected or significant. Loosely put the FDR may be interpreted as the proportion of false positives among those genes judged significantly regulated. Equation (2) is the FDR estimate presented in [3].
Denote by E[X] the expectation (or true mean) of any random variable X. With V the number of false positives given a certain cut-off and R the number of rejected null hypotheses, one may define the FDR as the expectation of the ratio of these quantities, or
where care is taken to avoid division by zero.
In [10] and [11] the FDR is estimated as the ratio of the expected proportion of false positives given the cut-off to the expected proportion selected. Viewed as a function of a cut-off α, such that genes g_{i} with p_{ i }less than α are judged significant in terms of p-values, following the continuous cumulative distribution function (cdf) F, the FDR estimate is
which is nearly equal to (2) with the exception that the P_{(L)}has been replaced by its upper bound, and the step-wise empirical distribution by a smooth version, either a parametric model or a smoothed version of the empirical distribution. Thus, the FDR is now a continuous function instead of piece-wise continuous with jumps at each observed p-value. This ratio of expected values tries to estimate the expectation of a ratio: In general such an approach will give an overestimation, but in practice this will have little effect, see the Additional file.
The related concept of the positive FDR, pFDR = E[V/R|R > 0], the expectation conditional on at least one rejection, appears in [2]. Other forms of FDR have been proposed such as the conditional FDR [2], cFDR, defined as the expected proportion of false positives conditional on the event that R = r rejected have been observed : cFDR(r) = E[V|R = r]/r. This would answer to the question "What proportion of false positives may I expect in my top list of r genes?". Under independence and identical distribution in a Bayesian setting it is proved in [23], that pFDR, cFDR and the marginal FDR, mFDR = E[V]/E[R] [2], all coincide with p_{0}α/F(α) at the cutpoint α, cf. (6).
Instead of p-values it has been suggested in to calculate q-values that verbally have the following meaning for an individual gene [2, 9]:
The q-value for a particular gene is the minimum false discovery rate that can be attained when calling all genes up through that one significant [9].
These q-values can be used to determine a cut-off similar to the classic 5% cut-off for univariate tests developed in statistics long ago. But in many applications one should not be too rigid about any particular value, since the emphasis often is on discovery rather than hypothesis testing: we generate hypotheses worthy of further investigation. Thus the balance between false positives and false negatives will be crucial: Rather than keeping the risk of erroneously selecting one individual gene at a fixed level, it is the decision involving thousands of genes given the amount of genes we can follow up on that is the focus, and where both types of error must be considered. The q-value does not fully address this problem, but nevertheless represents an improvement over the classical multiple test procedures in these applications.
More mathematically the q-value can be expressed as
Taking minimum in (7) enforces monotonicity in p_{ i }, so that the q-value will be increasing (non-decreasing) in the observed p-value. If the FDR is non-increasing, as it should, then q (p_{ i }) = FDR (p_{ i }).
Additionally, the FDR offers a framework for power and sample size calculations, see [24] and the new developments in [25].
Results
Eight estimates of p_{0} and six of FDR (based on six of the former) were tested on both simulated data and real data. The differing numbers are motivated below.
- 1.
the beta-uniform model (BUM) [10], which fits a mixture of a uniform and a beta distribution to the observed p-values; function ext.pi.
- 2.
spacing LOESS histogram (SPLOSH) [11], which fits a non-parametric spline that estimates the logarithm of the pdf;function splosh.
- 3.
- 4.
the smoother [9], which fits a spline to a function of a cut-off value, namely the proportion of p-values greater than that cut-off divided by the expected proportion under the uniform distribution;function qvalue.
- 5.
the bootstrap least squares estimate (bootstrap LSE) [3], which is related to the previous estimate;function qvalue or estimatep0.
- 6.
the Successive Elimination Procedure (SEP) [13];selects a subset which represents the null distribution by behaving like a uniform;function twilight.
- 7.
a new method based on a moment generating function approach (mgf);function p0.mom.
- 8.
The bootstrap estimate and mgf did not participate in the calculation of FDR. The smoother gives the basically same value as the bootstrap estimate, and mgf is unnecessarily conservative for lower values of p_{0}, compared to PRE.
- 1.
BUM FDR (based on BUM;function bum.FDR)
- 2.
BH FDR (based on LSL and function fdr.control).
- 3.
SPLOSH FDR (based SPLOSH;function splosh)
- 4.
smoother FDR or R function qvalue [9] (based on the smoother)
- 5.
SEP fdr (based on SEP;function twilight)
- 6.
the new method pava FDR (based on PRE;function pava.fdr)
For brevity mgf, PRE and pava FDR will all be referred to as new methods. It would be more exhaustive to say that PRE is a minor modification of an existing method [25] applied to p-values rather than test statistics and provided as a new implementation in R; and that pava fdr is based on [29] with local splines replaced by isotonic regression and provided as a new R function. On the other hand mgf seems quite new. More details follow in Methods.
For reference some graphs include an estimate of the SEP local FDR, defined as LFDR(p) = p_{0}/f(p), estimating the probability that a gene whose p-value equals p is a false positive. Furthermore, the ouput from R function locfdr applied to the real life data (with nulltype = 0, i.e. a standard normal distribution which has cdf Φ, see Methods) and the transformed t-test statistics : Z = Φ^{-1}(F(t)), and F the t-test distribution (details below) gives perspective on the other methods and opens up an alternative route to making inferences. This function produces an estimate of the local FDR as a function of the transformed test statistic Z [25]. In that same reference the author argues in favour of the cutpoint LFDR ≤ 0.2, which implies quite high posterior odds in favour of the non-null case : (1-p_{0})f_{1}/p_{0}f_{0} ≥ 4.
Simulated data
Two simulation models were used: one generating values independent between genes and the other generating observations displaying clumpy dependence [14, 30].
Simulation of independent data. Denote by N(μ, σ) a normal distribution with mean μ and standard deviation σ. For each DEG one of the above three scenarios was chosen with equal probabilities. For the rest both groups follow the same distribution chosen randomly from the 'Group 1' column. The scenarios are such that the power to detect the regulation with a 5% two-sided t-test ranges from small to large given two groups of size four.
Scenario | Group 1 | Group 2 | Power |
---|---|---|---|
1 | N(6, 0.1) | N(6.1, 0.1) | 0.19 |
2 | N(8, 0.2) | N(8.5, 0.2) | 0.79 |
3 | N(10, 0.4) | N(11, 0.7) | 0.47 |
To generate dependent data the protocol from [14] was followed. This generates data following clumpy dependence in the sense of [30] such that blocks of genes have dependent expression. First a logarithmic normal distribution with mean 1.5 and standard deviation 0.3 generated a profile for each gene. Denoting by N(μ, σ) a normal distribution with mean μ and standard deviation σ, random errors following a standard normal distribution N(0,1) were added. To create dependencies genes were partitioned into sets of 50 and for each sample the same term from a N(0, 1) distribution was added to the expression of each gene in the set. Finally, genes were randomly assigned to become DEGs with probability 1-p_{0} and for each gene a regulation term following either N(0.5, 0.2) or N(0.7, 0.2), with equal probabilities, was added to the expression of one of the groups of samples of size 30. The power to detect either of these two alternatives with a t-test at the 5% significance level equals 31% and 50%, respectively. The procedure generated for each of 400 iterations a set of observations of 10,000 genes. The protocol gives rise to high correlation within the blocks of 50 (on average on the order of 0.5). Results for smaller or weakly dependent datasets appear in the Additional file. Here weakly dependent means that the clumpy dependence term follows a N(0,1/20) distribution (correlation within blocks slumps to 0.003 on an average).
Over-all results of simulations of independent data. Data sets with p_{0} ranging from 0.6 to 0.99 were simulated. The summary statistics of the absolute difference between target value and its estimate show a rather varying performance for all methods, with PRE having the smallest bias and variation.
BUM | SPLOSH | smoother | bootstrap | SEP | LSL | mgf | PRE | |
---|---|---|---|---|---|---|---|---|
Mean | 0.039 | 0.061 | 0.038 | 0.036 | 0.045 | 0.18 | 0.072 | 0.022 |
Sd | 0.048 | 0.078 | 0.032 | 0.034 | 0.032 | 0.12 | 0.048 | 0.016 |
Detailed statistics on the estimates of p_{0} based on simulations of independent data. The displays the mean bias (true – estimated) and standard deviation of estimates for each level of true p_{0}.
True p _{0} | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | 0.95 | 0.99 | |
---|---|---|---|---|---|---|---|---|
BUM | mean bias | 0.013 | 0.043 | 0.038 | 0.028 | -0.090 | -0.050 | -0.010 |
Sd | 0.0044 | 0.0074 | 0.0072 | 0.0076 | 0.033 | 0.0021 | 0.0010 | |
SPLOSH | mean bias | -0.083 | -0.068 | -0.043 | -0.012 | 0.018 | 0.026 | 0.14 |
Sd | 0.014 | 0.020 | 0.028 | 0.034 | 0.040 | 0.041 | 0.063 | |
QVALUE | mean bias | -0.066 | -0.057 | -0.046 | -0.034 | -0.019 | -0.015 | 0.0012 |
Sd | 0.022 | 0.022 | 0.023 | 0.027 | 0.027 | 0.024 | 0.0160 | |
Bootstrap LSE | mean bias | -0.065 | -0.054 | -0.040 | -0.025 | -0.0067 | -0.0037 | 0.0057 |
Sd | 0.023 | 0.023 | 0.021 | 0.024 | 0.026 | 0.024 | 0.023 | |
SEP | mean bias | -0.084 | -0.072 | -0.059 | -0.043 | -0.028 | -0.020 | 0.0068 |
Sd | 0.014 | 0.013 | 0.014 | 0.013 | 0.013 | 0.012 | 0.014 | |
LSL | Mean bias | -0.36 | -0.31 | -0.25 | -0.18 | -0.096 | -0.049 | -0.010 |
Sd | 0.025 | 0.019 | 0.014 | 0.0086 | 0.0036 | 0.0022 | 0.0010 | |
mgf | mean bias | -0.15 | -0.12 | -0.095 | -0.067 | -0.040 | -0.027 | -0.0095 |
Sd | 0.0052 | 0.0052 | 0.0051 | 0.0056 | 0.0056 | 0.01098 | 0.0018 | |
PRE | mean bias | -0.036 | -0.033 | -0.028 | -0.018 | -0.0078 | -0.018 | -0.0018 |
Sd | 0.0098 | 0.010 | 0.0097 | 0.012 | 0.016 | 0.0080 | 0.0067 |
The new methods mgf and PRE were very competitive on these data, and had both low bias and variation, excluding mgf at the 0.5 and 0.6 level. Since mgf tends to overestimate p_{0} rather much in the lower range, one may prefer PRE. For practical purposes though overestimation is desirable and enables control of the error rate (exact control in the terminology from [31]).
The smoother and the bootstrap had good and quite similar performance. They give more or less the same variation and bias over the whole range. This variation can be a bit high though, especially when comparing to PRE.
In the higher range BUM gives a crude estimate of the true p_{0}. In a certain lower range however it underestimates. As we can see in Figure 1, however, the method considerably overestimates p_{0} in the higher range, which brings down the power to detect DEGs.
SPLOSH has the advantage of fitting the observed distribution quite well, judging from some tests (data not shown), compare also [11]. This enables a Bayesian analysis as in (3). However, it does the fitting of the p-values close to 0 sometimes at the expense of the accuracy concerning the values at the other end, thus misses the plateau and the minimisation in (10) will give a misleading result. In particular, this tends to happen when there are few DEGs. As we can see in Figure 1 the method underestimates p_{0} at the higher range (p_{0} ≥ 0.9), which is worrisome and may lead to underestimation of the error rate, which is undesirable for a method of statistical inference.
From Tables 2 and 3 we can see that PRE has the best over-all performance, followed by the smoother and the bootstrap. This does not however imply that the other methods could not be considered. The results vary quite a lot depending on the value of p_{0}: LSL is quite competitive for p_{0} = 0.99, but too conservative for p_{0} = 0.5.
Over-all results of simulations of dependent data. Data sets with p_{0} ranging from 0.6 to 0.99 were simulated. The summary statistics of the absolute difference between target value and its estimate show a rather varying performance for all methods, with BUM now having the smallest bias and variation with PRE in second place.
BUM | SPLOSH | Smoother | Bootstrap | SEP | LSL | Mgf | PRE | |
---|---|---|---|---|---|---|---|---|
Mean | 0.054 | 0.075 | 0.07.3 | 0.085 | 0.071 | 0.125 | 0.091 | 0.064 |
Sd | 0.035 | 0.086 | 0.062 | 0.076 | 0.062 | 0.124 | 0.069 | 0.060 |
Detailed statistics on the estimates of p_{0}based on simulations of dependent data. The table displays the mean bias (true – estimated) and standard deviation of estimates for each level of true p_{0}.
True p _{0} | 0.5 | 0.6 | 0.7 | 0.8 | 0.9 | 0.95 | 0.99 | |
---|---|---|---|---|---|---|---|---|
BUM | mean bias | -0.09885 | -0.0744 | -0.0524 | -0.0308 | -0.0525 | -0.0493 | -0.0100 |
Sd | 0.01555 | 0.0167 | 0.0169 | 0.0194 | 0.0520 | 0.0088 | 0.0010 | |
SPLOSH | mean bias | -0.15811 | -0.1179 | -0.0710 | -0.02570 | 0.0180 | 0.0318 | 0.06761 |
Sd | 0.02702 | 0.0309 | 0.0338 | 0.0398 | 0.0423 | 0.0379 | 0.0560 | |
QVALUE | mean bias | -0.15089 | -0.1217 | -0.0895 | -0.0593 | -0.0307 | -0.0142 | 0.0093 |
Sd | 0.02661 | 0.0306 | 0.0302 | 0.0345 | 0.0368 | 0.0309 | 0.0255 | |
Bootstrap LSE | mean bias | -0.18009 | -0.1499 | -0.1080 | -0.0691 | -0.0313 | -0.0126 | 0.0084 |
Sd | 0.04242 | 0.0382 | 0.0351 | 0.0326 | 0.0299 | 0.0269 | 0.0232 | |
SEP | mean bias | -0.15525 | -0.1238 | -0.0915 | -0.0584 | -0.0277 | -0.0134 | 0.0138 |
Sd | 0.02419 | 0.0239 | 0.0241 | 0.0265 | 0.0269 | 0.0203 | 0.0201 | |
LSL | Mean Bias | -0.2710 | -0.2203 | -0.1667 | -0.1110 | -0.0554 | -0.0292 | 0.0017 |
Sd | 0.13333 | 0.1119 | 0.0902 | 0.0671 | 0.0410 | 0.0221 | 0.0165 | |
Mgf | mean bias | -0.18322 | -0.1468 | -0.1349 | -0.0898 | -0.0460 | -0.0235 | 0.0007 |
Sd | 0.04633 | 0.0401 | 0.0140 | 0.0153 | 0.0160 | 0.0153 | 0.0128 | |
PRE | mean bias | 0.14088 | -0.1091 | -0.0770 | -0.0442 | -0.0152 | -0.0021 | 0.0194 |
Sd | 0.02300 | 0.0239 | 0.0239 | 0.0291 | 0.0310 | 0.0269 | 0.0248 |
Results for 300, 5,000 and 10,000 simulated weakly dependent genes appear in the Additional file. Briefly, they resemble those of the independent case. For 300 genes however the variation is such that the value of using these methods seems doubtful.
Real data
Data from Golub et al
Removing probe sets with less than 20% present rate will leave us with 2999 probe sets, and qvalue indicates that there are 977 of them that are significant with a q-value less than 5% (p_{0} = 0.47). In general it is wise to remove probe sets with low presence rate prior to analysis, since this will make the inference more reliable, compare [20]. Doing so will most likely produce more true positives.
Data from Spira et al
Next let us turn to a case where there are rather few DEGs. In [34] the results from a microarray experiment where bronchial epithelial brush biopsies have been hybridised to Affymetrix U133A arrays are presented. The biopsies come from three different subject categories: Current smokers, Never smokers and Former smokers. The cel intensity files were downloaded from the NCBI Gene Expression Omnibus (accession no. GSE994) [35]. The Bioconductor package affy [19] was used to normalise intensities with the quantile method, and to calculate the RMA measure of abundance. The function mas5calls in affy output absolute calls.
Here we will take a brief look at the comparison between Former smokers and Never smokers. The comparison may help identify genes that remain changed after smoke cessation. A fuller analysis would include more analyses, such as the Current smokers vs. Former smokers comparison, and possibly also adjust for the fact that Former smokers tend to be older than Never smokers (Mean Age 45 and 53 years, respectively).
Figure 14 displays the graphical output from locfdr, where the LFDR seems to approach zero in the tails, contradicting SEP, SPLOSH and BUM, but essentially agreeing with the concordant trio BH, pava and qvalue.
Discussion
Over-all results will lump together performance under different conditions and may thus be less relevant for a particular application. For instance, in practice the performance for high p_{0} will probably matter more than that for lower values. When many genes are changed the cutpoint will likely be chosen based on other criteria than FDR, and hence the difference between methods becomes less relevant. However, the detailed results presented here should give the practitioner some guidance as to what methods could be considered. Looking at the p-value histogram one can find some decision support in the choice of method. Comparing the output from several methods provides further clues.
All the methods performed worse on the dependent data; both the estimate of p_{0} and FDR suffered. To some extent that may be due to the lower mean power of the alternatives in that simulation model. However, the methods were derived under the assumption of independence and the small difference in mean power of 0.08 does not explain the great deterioration in most methods. Indeed, for simulated datasets with weak dependencies the results came close to the independent case, see the Additional file.
Through all tests PRE and pava FDR proved quite successful. Of the methods for estimating FDR, qvalue has the advantage of a well-documented and good track record, and behaves well here. BUM displays a varying performance, but does handle dependent data well. In practice it will be difficult to know where on the scale from independence to strong clumpy dependence a particular dataset will rate, if indeed it follows clumpy dependence at all. LSL, and BH, have some problems, but on the other hand they arise mainly at low p_{0}, where they probably matter the least. As noted above regarding the Spira et al. data, SEP LFDR and FDR stabilised at a value above 0.8 when the cut-off approaches zero. In other tests SEP performed well, particularly with independent data. BUM in this case produced the estimate FDR ≡ 1 which can hardly reflect the truth.
The locfdr method offers the possibility to choose between three different null type distributions. The choice of the null type N(0,1) produced p_{0} estimates similar to those of the other methods. The need to specify the transform m may seem like an obstacle. But in many situations a parametric test statistic with a known null distribution exists. Alternatively, m could be identified by modelling a bootstrap distribution [25].
All the described methods assume the p-values were obtained in a reliable fashion, e.g. by a warranted normal approximation, a bootstrap or a permutation method. Reference [10] describes a case when a two-way ANOVA F-distribution was used when the distributional assumptions were not met. The estimate of p_{0} gave an unrealistic answer. When permutation p-values were used instead their method gave a more realistic result. One always has to bear this caveat in mind. To further complicate matters, the permutation of sample labels approach is no panacea if the independence between samples assumption does not hold true, as detailed in [27]. (Let us follow the usual convention that genes come in rows of the data matrix, and samples in columns.) Permuting within columns provided some remedy there. Misspecifying the null distribution will jeopardize any simultaneous inference, whether based on FDR or not. It may pay off to consider the correlation structure in data, both in view of this finding and in view of the different performance of methods depending on the strength of correlations.
The q-value q(p_{ i }) has been criticised for being too optimistic in that it weighs in also genes that are more extreme than i when calculating the measure. Note that a similar criticism could be levied against the classical p-value: the p-value gives the probability under the null hypothesis of observing a test statistic at least as extreme as the one observed. Also, there is no clear stable, reliable and tested alternative. This is not to say that the q-value is unproblematic, but it still has been studied and used much more than e.g. the local FDR, which may suffer from high random variation, see examples in [29]. Other examples from Results section give evidence of stability. Contrary to what one may anticipate the FDR is not always more stable than LFDR [25]. The concept of a local FDR seems quite interesting and may lead the way towards improved inference, and it begs a thorough investigation of the various recently published options.
To avoid pit-falls in the inference one must use the total information obtained from p_{0} and the FDR or q-value curve, see also Storey in the discussion of [31]. There is not one cut-off in terms of q-value that will suit all problems. Take the case of Figure 12, where one will have to accept a high FDR in order to find any DEGs. At the other end of the spectre, in Figure 7, the cut-off can be much more restrictive. The choice of cut-off must be made with a view to one's belief regarding p_{0}, and calculating the sum of (4) and (5) to assess to total of false positives and false negatives gives further guidance in this choice. In general it makes sense to choose a cut-off in the region [0, α_{ min }], where α_{ min }is the value which minimises the total relative frequency of errors committed FP(α)+FN(α), see (4) and (5). However, since false positives and false negatives have different consequences with possibly different losses, it is difficult to state an algorithm that would cover all scenarios.
Conclusion
This article deals in the main with a simple frequentist framework for the analysis of microarray experiments. The conclusion is that the concept of the proportion of unchanged genes and the related concept of a q-value or false discovery rate are practical for such analysis. Furthermore, there exists open source code that implements methods that address the needs of the practitioner in this field. New methods gave evidence of improved performance, allowing better control of the error rate and thus enabling a more careful identification of DEGs. Issues still remain and improvements will probably appear over the next couple of years, but as a provisional solution these methods have much to offer.
Methods
The current article focuses on the two-component model. Other points of view exist. In reference [25] the two-component model is reshaped into a conceptually attractive one-group model allowing a continuum of effects.
Denote the pdf of p-values by f, the proportion of unchanged by p_{0} and the distribution of the p-values for the changed genes by f_{1}. Then the pdf of p-values may be written as
f(x) = p_{0} × 1 + (1 - p_{0})f_{1}(x) (8)
using the fact that p-values for the unchanged genes follow a uniform distribution over the interval [0,1]. This model is unidentifiable without further assumptions, e.g. that p-values in a vicinity of 1 only represent unchanged genes. From the non-negativity of pdf's, clearly
f(x) ≥ p_{0} (9)
This leads to the estimate based on the minimum of the estimated pdf [1]
see also Figure 6. In most cases the minimum in (10) will occur for some x close to or at 1. Hence (10) will in these cases agree well with an estimate of f(1). If one has reason to believe that p_{0} is close to 1, it may pay off to replace (10) by the 25% percentile or simply put the estimate equal to 1, in order to make overestimation more likely.
LSL
Let R(α) = # {i : p_{ i }≤ α}, the number of rejected given the cut-off α. In [36] the approximation
N - R(α) ≈ E[N - R(α)] ≈ N_{0}(1 - α)
for small α and N_{0} = Np_{0} the number of true null hypotheses appears. Consequently, (N - R(p_{(i)})/(1 - p_{(i)}) = (N - i)/(1 - p_{(i)}) will approximate N_{0}, which lead the pioneering authors to consider plotting 1 - p_{(i)}against N - i, thus giving them an estimate of N_{0}. In [26] the Lowest SLope estimator (LSL) of N_{0} based on the slopes S_{ i }= (1 - p_{(i)})/(N - i + 1) is presented. Starting from i = 1, the procedure stops at the first occurrence of Si_{0} <Si_{0}-1, and outputs the estimate
In [12] the two above estimates are presented, derived and compared together with a method called Mean of Differences Method (MD). MD and LSL are motivated by assuming independence and approximating the gaps d_{(i)}= p_{(i)}- p_{(i-1)}(define p_{(0)} = 0 and p_{(N+1))}≡ 1) with a Beta(1, N_{0}) distribution, which has expectation 1/(N_{0} + 1). This expectation may be estimated by the inverse of a mean of the form
MD proceeds downward and chooses i_{0} equal to the first j satisfying
Of these three methods LSL and MD give very similar results, and outperform their predecessor [12].
LSL is available as function fdr.estimate.eta0 in package GeneTS [18] with the option method= "adaptive".
The smoother
A method here referred to as the smoother appeared in [9]. This method, like all presented, is based on a comparison of the empirical p-value distribution to that of the uniform distribution. There will likely be fewer p-values close to 1 in the empirical than in the null distribution, which is a uniform. The ratio of the proportion of p-values greater than some η to the expected proportion under the uniform distribution, 1-η, will give a measure of the thinning of observed p-values compared to the null distribution. Thus, with F_{ e }denoting the empirical distribution, the ratio {1-F_{ e }(η)}/{1-η} will often be a good estimate of p_{0} for an astutely chosen threshold η. A spline is fitted to the function p_{0}(η) = {1-F_{ e }(η)}/{1-η}, and the resulting function is evaluated at η = 1, yielding the estimate
The q-value is estimated by combining (2), (7) and (11), and an implementation is provided as the function qvalue in package qvalue available on CRAN [18].
In [30] the authors go to great lengths to prove that for fixed η, as above, the conservativeness remains under various forms of dependency, such as clumpy dependence.
The Bootstrap LSE
Another approach pioneered by Storey in [3] is to use a bootstrap least squares estimate (LSE), which chooses a value of η in p_{0}(η), that minimises the variation of the estimate for random samples of the original p-values. The bootstrap standard reference [37] provides more theoretical background. Generate B new samples p-values p*^{ b }(b = 1,..., B) by sampling with replacement from the observed ones, calculate a measure of the Mean Squared Error (MSE) of the corresponding estimates p*^{ b }_{0}(η) for a lattice of values of η and choose the value minimising the MSE. More formally, the optimal η is obtained through
The version of the bootstrap used in this article uses more samples B than the version available in qvalue (B = 500 instead of B = 100), and seems to perform better (data not shown).
Available in functions qvalue [18] and p0.mom (in package SAGx) [18, 38].
SPLOSH
In [11] a spline function estimates the log-transformed pdf log[f(x)] using a complex algorithm involving splines called spacings LOESS histogram (SPLOSH). To obtain a stable estimate of FDR near zero a technique from mathematical analysis called l'Hospital's rule is used to approximate the ratio in (5) and to yield
where the numerator has been estimated as in (4). An R package with the same name is available [39].
The FDR estimate (6) is used with F obtained by the non-parametric estimate of the pdf.
Note that we can now calculate the posterior probability given its p-value that a gene is a DEG as p_{1}(x) = 1 - p_{0}/f(x), compare (3).
The method is available in R function splosh [40].
BUM
In [10] the authors assume a beta-uniform (BUM) distribution, i.e. in (1) they replace f_{1} by a beta distribution,
f(x) = λ + (1 - λ)ax^{a-1} (12)
where in addition to λ which corresponds to p_{0} the shape parameter a has to be estimated. Thanks to the simple form of the distribution it is possible to estimate parameters through the maximum likelihood principle, i.e. by choosing values that maximise the likelihood of observing the p-values that were actually observed. However, due to problem in identifying p_{0} with λ, the authors instead use an upper bound
which corresponds to f(1).
The FDR estimate (5) is used with F the cdf corresponding to (12).
The authors provide R code for the application of their method [39].
A more intricate Hierarchical Bayes model based on the beta-uniform concept allowing for different parameter values in different intervals appears in [41]. The R function localFDR provides an implementation of the method [42].
Poisson regression
In [28] it is suggested to estimate any empirical distribution by dividing the real axis into intervals and regarding the number of hits in each interval as the result of an inhomogeneous Poisson process, much like counting the number of cars arriving at a crossing during different time intervals. This method was used in [27] to model the distribution of a transformed test statistic, it also appears in function locfdr which estimates a local FDR as a function of a test statistic. In our case, of course, the support of the distribution is the unit interval [0,1]. Then the expected number of hits in each subinterval of [0,1] can be modelled as a polynomial in the midpoints of the subintervals by a technique called Poisson regression (PRE). The approach taken here is to choose a polynomial of low degree so that the plateau representing the uniform distribution is well captured. In doing so the ability the capture the distribution at low p-values is sacrificed.
A more mathematical description now follows. The PRE method assumes that the counts S_{ k }follow a Poisson distribution whose intensity is determined by the midpoint t_{ k }of the interval I_{ k }, see [28]. To be specific: in the current application it is assumed that the expected frequency of observations in an interval is given by
where μ^{ o }_{ k }are the smoothed observed frequencies in each interval I_{ k }. In statistical jargon this is a Poisson regression model with μ^{ o }_{ k }as offset. This assumes independence between counts in different intervals. Although this does not hold true the model succeeds to capture the essential features of distributions. Standard functions in e.g. R can fit this model. Normalising the curve by the total number of p-values we get an estimate of the pdf. Finally, smooth the pdf f(x) with a spline to obtain a more stable result, and use the estimate (10). An implementation of PRE is provided through R function p0.mom in package SAGx [18, 38].
SEP
The Successive Elimination Procedure (SEP) excludes and includes p_{ i }successively such that the similarity of the distribution of the included tends to behave increasingly like a uniform [13]. Finally, an index set J_{ final }will map to a set of p-values that represent the true null hypotheses. This yields the point estimate
with N_{ J }= # J, the cardinality of the set J. The identification of J_{ final }proceeds by an intricate optimisation algorithm where the objective function consists of two terms : one Kolmogorov-Smirnov score
for the empirical cdf F_{ J }(based on J), to measure the distance to a uniform, and one penalty term to guard against overfitting
for some tuning parameter λ.
A local FDR is obtained from smoothed histogram estimates based on equidistant bins
where the function h_{0} refers to the J_{ final }set and h to the total set of p-values.
The function twilight in package twilight provides an implementation of SEP [19].
Moment generating function approach
The next approach is based on the moment generating function (mgf), which is a transform of a random distribution, which yields a function M(s) characteristic of the distribution, cf. Fourier or Laplace transforms, e.g. [43]. Knowing the transform means knowing the distribution. It is defined as the expectation (or the true mean) of the antilog transform of s times a random variable X, i.e. the expectation of e^{ sX }or in mathematical notation:
M(s) = ∫e^{ sx }f(x)dx.
To calculate the mgf for p-values, we use the fact that the pdf is a mixture of pdf's (8). This yields the weighted sum of two transformed distributions:
where we have used the fact that the mgf of a uniform distribution over [0,1] equals g(s) = (e^{ s }- 1)/s. Denoting the second transform by M_{1}(s) we finally have
M(s) = p_{0}g(s) + (1 - p_{0})M_{1}(s). (13)
Now, the idea is to estimate these mgf's and to solve for p_{0}. In the above equation M(s) can be estimated based on an observed vector of p-values and g(s) can be calculated exactly, respectively, while p_{0} and M_{1}(s) cannot be estimated independently. The estimable transform is, given the observed p-values p = p_{1},..., p_{ n }, estimated by
Then, one can solve (13) for p_{0}:
Let us do so for s_{ n }> s_{n-1}, equate the two ratios defined by the right hand side in (14) and solve for M_{1}(s_{ n }). This gives the recursion
where s = (s_{1}, ..., s_{ n }). The CV will in contrast to the variance put both small and high values of the ratios on an equal footing and enable comparison.
Finally, these essentially equal ratios provide an estimate of p_{0}.
A heuristic convexity argument suggests that mgf over-estimates p_{0}, see the Additional file. Furthermore, the bias seems to decrease as p_{0} grows.
An implementation of mgf appears as function p0.mom in package SAGx [18, 38].
Local FDR and FDR
The concept of a local false discovery rate originates from [1]. Let the (true) local FDR at t be defined as the probability that gene i is unchanged conditional upon that its p-value equals t, or in formulas : LFDR(t) = Pr(gene i unchanged | p_{ i }= t) = p_{0}/f(t). The Averaging Theorem of [4] states that integrating the local FDR over the rejection region R, such as R = [0, 0.01], yields the FDR : FDR(R) = E[LFDR(y) | y ∈ R]. In [29] it is noted that the estimated q-value equals the mean of a local FDR
where the local FDR at the i^{ th }ordered p-value p_{(i)}equals
where N denotes the total number of genes. This rephrases the theorem in terms of estimators. The local FDR is meant as an approximation of the probability that an individual gene i is a DEG. As remarked in [29] the q-value does not estimate the probability that a gene is a false positive. Indeed, the theorem shows that it is the mean of that probability for all genes at least as extreme as i. Thus the q-value will tend to give a lower value than LFDR(i).
Under a wide range of models, where f(x) is non-increasing, e.g. the BUM model, the expected local LFDR(i) will be non-increasing, and hence the differences above should tend to increase, see the Additional file. Hence there is a need for enforcing monotonicity as in (7). One tool for enforcing monotonicity is the Pooling of Adjacent Violators (PAVA) algorithm [44]. This algorithm has an intuitive appeal, is less ad-hoc than the local spline approach presented in [29], and is the non-parametric maximum likelihood estimate under the assumption of monotonicity. As an example of how it works, consider the series (1,2,4,3,5), which PAVA turns into the non-decreasing series (1, 2, 3.5, 3.5, 5) by pooling the violators of monotonicity (4, 3) and replacing them by their mean. Though not equivalent to the q-value from (2) and (7), the results from applying PAVA to the terms in (15) agreed rather well with the values obtained from function qvalue. In the Results section this approach combined with the PRE estimate of p_{0} is referred to as pava FDR. We could have used mgf for calculating FDR, but it was excluded due to the better over-all performance of PRE.
The bootstrap LSE gives a very similar result to the smoother and thus was excluded in comparison of FDR estimates.
The R function pava.fdr in package SAGx provides an implementation of pava FDR, and returns a list of estimates of FDR, LFDR and p_{0} [18, 38].
Declarations
Acknowledgements
Thanks are due to Niclas Sjögren at AstraZeneca R&D Södertälje for valuable comments. Furthermore, suggestions made by the three reviewers improved both content and presentation.
Authors’ Affiliations
References
- Efron B, Tibshirani R, Storey JD, Tusher VG: Empirical Bayes analysis of a microarray experiment. Journal of the American Statistical Association 2001, 96: 1151–1160. 10.1198/016214501753382129View ArticleGoogle Scholar
- Benjamini Y, Hochberg Y: Controlling the false discovery rate : a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society B 1995, 57: 289–300.Google Scholar
- Storey JD: A Direct Approach to False Discovery Rates. Journal of the Royal Statistical Society B 2001, 64: 479–498. 10.1111/1467-9868.00346View ArticleGoogle Scholar
- Efron B, Storey JD, Tibshirani R: Microarrays, Empirical Bayes Methods, and False Discovery Rates. In Technical report. Volume 217. Stanford, Department of Statistics, Stanford University; 2001.Google Scholar
- Broberg P: Statistical methods for ranking differentially expressed genes. Genome Biology 2003, 4: R41. 10.1186/gb-2003-4-6-r41PubMed CentralView ArticlePubMedGoogle Scholar
- Delongchamp RR, Bowyer JF, Chen JJ, Kodell RL: Multiple-testing strategy for analyzing cDNA array data on gene expression. Biometrics 2004, 60: 774–782. 10.1111/j.0006-341X.2004.00228.xView ArticlePubMedGoogle Scholar
- Tusher VG, Tibshirani R, Chu G: Significance analysis of microarrays applied to the ionizing radiation response. PNAS 2001, 98: 5116–5121. 10.1073/pnas.091062498PubMed CentralView ArticlePubMedGoogle Scholar
- Allison DB, Gadbury GL, Moonseong H, Fernandez JR, Cheol-Koo L, Prolla TA, Weindruch RA: A mixture model approach for the analysis of microarray gene expression data. Computational Statistics and Data Analysis 2002, 39: 1–20. 10.1016/S0167-9473(01)00046-9View ArticleGoogle Scholar
- Storey JD, Tibshirani R: Statistical significance for genomewide studies. PNAS 2003, 100: 9440–9445. 10.1073/pnas.1530509100PubMed CentralView ArticlePubMedGoogle Scholar
- Pounds S, Morris SW: Estimating the occurrence of false positives and false negatives in microarray studies by approximating and partitioning the empirical distribution of p-values. Bioinformatics 2003, 19: 1236–1242. 10.1093/bioinformatics/btg148View ArticlePubMedGoogle Scholar
- Pounds S, Cheng C: Improving false discovery rate estimation. Bioinformatics 2004, 20: 1737–1745. 10.1093/bioinformatics/bth160View ArticlePubMedGoogle Scholar
- Hsueh H, Chen JJ, Kodell RL: Comparison of Methods for Estimating the Number of True Null Hypotheses in Multiplicity Testing. Journal of Biopharmaceutical Statistics 2003, 13: 675–689. 10.1081/BIP-120024202View ArticlePubMedGoogle Scholar
- Scheid S, Spang R: A Stochastic Downhill Search Algorithm for Estimating the Local False Discovery Rate. IEEE Transactions on Computational Biology and Bioinformatics 2004, 1: 98–108. 10.1109/TCBB.2004.24View ArticlePubMedGoogle Scholar
- Scheid S, Spang R: A False Discovery Approach to Separate the Score Distributions of Induced and Non-induced Genes. In Proceedings of the 3rd International Workshop on Distributed Statistical Computing (DSC 2003). Edited by: Hornik K, Leisch F and Zeilis A. Vienna, Austria, Technische Universität Wien. Institut für Statistik; 2003.Google Scholar
- Lehmann EL: Testing Statistical Hypotheses. New York, Wiley; 1953.Google Scholar
- R Foundation for Statistical Computing: R Development Core Team. R: A language and environment for statistical computing.[http://www.R-project.org]
- Ihaka R, Gentleman R: R: A language for data analysis and graphics. Journal of Computational and Graphical Statistics 1996, 5: 299–314.Google Scholar
- The Comprehensive R Archive Network[http://www.cran.r-project.org]
- Bioconductor[http://www.bioconductor.org]
- Yang YH, Speed T: Design and analysis of comparative microarray experiments. In Statistical Analysis of Gene Expression Microarray Data. Edited by: Speed T. London, Chapman & Hall/CRC; 2003.Google Scholar
- Simes R: An improved Bonferroni Procedure for Multiple Tests of Significance. Biometrika 1986, 73: 751–754.View ArticleGoogle Scholar
- Reiner A, Yekutieli D, Benjamini Y: Identifying differentially expressed genes using false discovery rate controlling procedures. Bioinformatics 2003, 19: 368–375. 10.1093/bioinformatics/btf877View ArticlePubMedGoogle Scholar
- Tsai CA, Hsueh H, Chen JJ: Estimation of False Discovery Rates in Multiple Testing: Application to Gene Microarray Data. Biometrics 2003, 59: 1071–1081. 10.1111/j.0006-341X.2003.00123.xView ArticlePubMedGoogle Scholar
- Yang MCK, Yang JJ, McIndoe RA, She JX: Microarray experimental design: power and sample size considerations. Physiol Genomics 2003, 16: 24–28. 10.1152/physiolgenomics.00037.2003View ArticlePubMedGoogle Scholar
- Efron B: Selection and Estimation for Large-Scale Simultaneous Inference. In Technical report. , Department of Statistics, Stanford University; 2004.Google Scholar
- Benjamini Y, Hochberg Y: On the adaptive control of the false discovery rate in multiple testing with independent statistics. Journal of Educational and Behavioural Statistics 2000, 25: 60–83.View ArticleGoogle Scholar
- Efron B: Large-scale simultaneous hypothesis testing : the choice of a null hypothesis. Journal of the American Statistical Association 2004, 99: 96–104.View ArticleGoogle Scholar
- Efron B, Tibshirani R: Using specially designed exponential families for density estimation. The Annals of Statistics 1996, 24: 2431–2461. 10.1214/aos/1032181161View ArticleGoogle Scholar
- Aubert J, Bar-Hen A, Daudin JJ, Robin S: Determination of the differentially expressed genes in microarray experiments using local FDR. BMC Bioinformatics 2004, 5: 125. 10.1186/1471-2105-5-125PubMed CentralView ArticlePubMedGoogle Scholar
- Storey JD, Tibshirani R: Estimating false discovery rates under dependence, with applications to DNA microarrays. Technical Report 2001–28, Department of Statistics, Stanford University 2001.Google Scholar
- Ge YC, Dudoit S, Speed TP: Resampling based multiple testing for microarray data analysis. Test 2003, 12: 1–77.View ArticleGoogle Scholar
- Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP, Coller H, Loh ML, Downing JR, Caligiuri MA, Bloomfield CD, Lander ES: Molecular Classification of Cancer: Class Discovery and Class Prediction by Gene Expression Monitoring. Science 1999, 286: 531–537. 10.1126/science.286.5439.531View ArticlePubMedGoogle Scholar
- Affymetrix web site[http://www.affymetrix.com]
- Spira A, Beane J, Shah V, Liu G, Schembri F, Yang X, Palma J, Brody JS: Effects of cigarette smoke on the human airway epithelial cell transcriptome. PNAS 2004, 101: 10143–10148. 10.1073/pnas.0401422101PubMed CentralView ArticlePubMedGoogle Scholar
- Gene Expression Omnibus[http://www.ncbi.nlm.nih.gov/geo]
- Schweder T, Spjotvoll E: Plots of p-values to evaluate many tests simultaneously. Biometrika 1982, 69: 493–502.View ArticleGoogle Scholar
- Efron B, Tibshirani R: An Introduction to the Bootstrap. London, Chapman & Hall; 1993.View ArticleGoogle Scholar
- The SAG homepage[http://home.swipnet.se/pibroberg/expression_hemsida1.html]
- BUM software[http://www.stjuderesearch.org/depts/biostats/BUM/index.html]
- SPLOSH web page[http://www.stjuderesearch.org/depts/biostats/documents/splosh-code3.ssc]
- Liao JG, Lin Y, Selvanayagam ZE, Shih WJ: A mixture model for estimating the local false discovery rate in DNA microarray analysis. Bioinformatics 2004, 20: 2694–2701. 10.1093/bioinformatics/bth310View ArticlePubMedGoogle Scholar
- localFDR web site[http://www.geocities.com/jg_liao]
- Feller W: An Introduction to Probability Theory and Its Applications. Volume 2. Second edition. New York, John Wiley; 1971.Google Scholar
- Ayer M, Brunk HD, Ewing GM, Reid WT, Silverman E: An empirical distribution function for sampling with incomplete information. Annals of Mathamatical Statistics 1955, 26: 641–647.View ArticleGoogle 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.