Abstract
Background
For RNAseq data, the aggregated counts of the short reads from the same gene is used to approximate the gene expression level. The count data can be modelled as samples from Poisson distributions with possible different parameters. To detect differentially expressed genes under two situations, statistical methods for detecting the difference of two Poisson means are used. When the expression level of a gene is low, i.e., the number of count is small, it is usually more difficult to detect the mean differences, and therefore statistical methods which are more powerful for low expression level are particularly desirable. In statistical literature, several methods have been proposed to compare two Poisson means (rates). In this paper, we compare these methods by using simulated and real RNAseq data.
Results
Through simulation study and real data analysis, we find that the Wald test with the data being logtransformed is more powerful than other methods, including the likelihood ratio test, which has similar power as the variance stabilizing transformation test; both are more powerful than the conditional exact test and Fisher exact test.
Conclusions
When the count data in RNAseq can be reasonably modelled as Poisson distribution, the WaldLog test is more powerful and should be used to detect the differentially expressed genes.
Background
Recent advancements in deep sequencing technique enable the ultrahighthroughput sequencing (called the second or next generation sequencing) approaches, to be used for transcriptome, including gene expression, analyses [14]. RNAseq produces a count (the total number of short reads which are annotated to positions within a gene or transcript) for each gene expression level. Some researchers have shown that the count data can be quite reasonably modelled as Poisson distribution [1,3]. To detect differentially expressed genes (DEGs) under two conditions (e.g., disease vs. healthy status), some statistical methods have been used. For example, Marioni et al [3] have proposed to use likelihood ratio test (LRT) based on a generalized linear model (GLM) with dependent variable (count) having Poisson distribution to detect DEGs under two conditions without biological replicates. Fisher exact test has also been proposed to detect DEGs for RNAseq data [5,6]. Some researchers modelled the count data as binomial distribution bin(n,p), where n is the total counts on the same lane, which usually ranges from several millions to tens of millions [7]. However, it is well known that for a binomial distribution bin(n,p) with large n and small p, it can be approximated by Poisson distribution. It is expected that for RNAseq data methods based on binomial distribution should have similar performance as those methods based on Poisson distribution and therefore are not considered in this paper.
If there are biological replicates, the count data may have larger variances than expected from the Poisson distribution as biological replicates bring extra variances. Under this kind of situations, Poisson distribution with overdispersed variances or negative binomial distributions are thought to be more appropriate [8,9]. However, due to the high cost, recent studies rarely used biological replicates; we therefore focus on the situations where at most technical replicates are used and Poisson distribution is assumed to be appropriate.
Besides Fisher exact test and LRT from GLM, several other test approaches in statistical literature have been proposed to detect the difference of two Poisson means. For example, under the null hypothesis that X_{1 }and X_{2 }are both from Poi(λ), then given k = X_{1}+X_{2}, X_{1 }has a binomial distribution Bin(k,p), where p = X_{1}/k. Therefore the conditional exact test can be used to detect the difference of the two Poisson means. One of the advantages of the GLM is that it can incorporate some covariates that we are interested in. If no covariates are considered due to the small sample size and large number of genes, the LRT can be constructed directly. Similar to LRT, Wald test also has an asymptotical Chisquare distribution with df = 1 [1012]. A modified Wald test (WaldLog) with data being first logtransformed has also been proposed for detecting the difference between two Poisson means [11]. Another method called variance stabilizing transformation (VST) was proposed by Huffman [10,13].
In this paper, we first compare the above mentioned methods by using simulations. Since for those methods having the same asymptotic distribution, they are expected to have similar performances for large sample sizes (e.g., Poisson distributions with large means), we focus on simulating Poisson distributions with small or moderate means. We then use a real RNAseq data to show the performances of those methods.
Results
Simulation results
In the simulations, we first assume there are no replicates, i.e., each of the two conditions has only one sample. We assume the data for condition 1 is from a Poisson distribution with fixed mean 5, 10, 15, or 30. For condition 2, we assume the Poisson mean is the same as or greater than that for condition 1. We use the nominal significance level 10^{3 }to reflect the situation where in a study with a large number of significant variable (genes), a stringent pvalue is needed. The estimated size and power are calculated from 10,000 runs for each setting.
Table 1 reports the estimated size and power for different situations. Figure 1 plots the estimated size and power for each method from this simulation.
Table 1. Estimated size and power (no replicates for each condition).
Figure 1. Estimated power from various methods (no replicates). Estimated size and power from each method for detecting the difference of two Poisson means. There are no replicates for each condition. The Poisson means for the first condition are 5, 10, 15, or 30. The second Poisson means are the same as (for size) or larger than (for power) the first ones. The nominal size (significance level) is 10^{3 }and 10,000 runs are used for each setting.
In the second simulation, we assume there are three replicates for each condition. Table 2 shows the estimated sizes and powers and these results are plotted in Figure 2.
Table 2. Estimated size and power (each condition has three replicates).
Figure 2. Estimated power from various methods (with 3 replicates for each condition). Estimated size and power from each method for detecting the difference of two Poisson means. There are three replicates for each condition. The Poisson means for the first condition are 5, 10, 15, or 30. The second Poisson means are the same as (for size) or larger than (for power) the first ones. The nominal size (significance level) is 10^{3 }and 10,000 runs are used for each setting.
From the results of the two simulation studies, we can see that the WaldLog method is usually more powerful than the LRT, which has similar power as that of VST; both are slightly better than the Wald test, which in turn outperforms Fisher test and conditional binomial test. It is noticeable that both Fisher and conditional tests are exact conditional tests and have almost identical powers.
Detecting DEGs for a real RNAseq dataset
We also apply these statistical methods to a real RNAseq data generated by Illumina's sequencing technology [3]. In this dataset, we consider two different samples: kidney and liver both at concentration 3PM, each has 5 lanes (technical replicates). We want to detect DEGs between these two samples. There are 15227 genes with at least 5 counts in total (i.e., on average 1 count on each lane) for both samples. For each of the 15227 genes, we calculate the pvalues using various methods mentioned above. Due to the almost identical performance between the Fisher and conditional binomial tests, we choose not to use the latter one in this real data analysis. Figure 3 plots the pvalues from other methods compared with the LRT. For many genes, WaldLog produces smaller pvalues than the LRT does.
Figure 3. Pvalues for the real RNAseq data from different methods. Compare pvalues (log10 scale) obtained from other methods with those from LRT for the 15227 genes whose average counts per lane are at least one for each sample.
We are particularly interested in seeing how these statistical methods perform for low expression data. To this purpose, we use the data from genes with average count per lane between 1 and 10. There are 2657 such genes from the data. Figure 4 plots the pvalues (on log10 scale) obtained from other methods compared with those from LRT for those 2657 genes. LRT, Fisher and VST obtain very similar pvalues for those genes, while Wald produces larger pvalues than LRT for a small portion of those genes. However, for many genes with small pvalues from LRT, their pvalues obtained by Waldlog are smaller. This indicates the WaldLog test is more powerful than LRT and other tests for this data.
Figure 4. Pvalues of the 2657 low espression genes from different methods. Compare pvalues (log10 scale) obtained from other methods with those from LRT for the 2657 low expression genes whose average counts per lane are between 1 and 10 for each sample.
To see how statistical significance (pvalues) related to biological significance (fold changes), we plot the pvalues (on log10 scale) vs. absolute fold change (on log2 scale) in Figure 5 from different methods for those 2657 genes. It can be seen that the WaldLog test obtains smaller pvalues than other methods when the fold changes are large, indicating this method can detecting more biologically meaningful DEGs than other methods for low expression genes.
Figure 5. Pvalues vs. log2(FC) of the 2657 low espression genes from different methods. Pvalue (log10 scale) obtained by various methods vs. fold change (log2 scale) for the 2657 low expression genes whose average counts per lane are between 1 and 10 for each sample.
Figure 6 plots the counts under two conditions for each low expression genes selected by WaldLog and LRT at significance level 10^{3}. Of the 581 selected genes by WaldLog, 498 are also selected by LRT. Each of the selected gene has at least 20 ounts in total for at least one sample (liver or kidney). The fold changes (liver vs. kidney, or kidney vs. liver) of all the 581 selected genes are all at least 2. For those 83 genes selected by WaldLog but not by LRT, some even have fold changes as large as 4.
Figure 6. Counts under two conditions for each of the selected low expression genes. With significance level 10^{3}, the WaldLog test selects 581 out of 2657 low expression genes. While the LRT selects 498 of the 581 genes. The count for each selected gene of the two samples is plotted.
Table 3 lists the numbers of DEGs obtained with different cutoff values by various methods. Consistent with what we have observed, the WaldLog test identifies more DEGs than other methods for low expression genes. This is also true when we apply these statistical tests to all of the 15227 genes.
Table 3. Numbers of DEGs of low expression obtained by various methods.
Discussion
Some studies have shown that for RNAseq data without biological replicates, the count data can be reasonably modelled as Poisson distributed [1,3]. Since it is usually more difficult to detect DEGs with low expression level for RNAseq data [14], it is desirable to find powerful statistical tests for detecting low expression DEGs. Through simulations and real data, we show that the WaldLog is more powerful that the commonly used LRT and Fisher exact test. The test statistic of WaldLog has an asymptotic normal distribution. Like many other test statistics having asymptotic normal or Chisquare distributions, the asymptotic distributional results for the WaldLog test statistic do not hold when the sample sizes are small (i.e., the Poisson mean is small). However, from our simulations we can see that when the Poisson mean is greater than 10, it can have reasonable sizes even if there are no replicates.
It is interesting to see that the WaldLog test is also related to the fold changes (see method section for more details); therefore it is also more biologically meaningful. The test statistic of WaldLog not only uses the fold change (on log scale), but also considers the variances associated with FC. For low expression level, the FC has larger variances, indicating it is more difficult to identify DEGs for low expression genes. This is consistent with the observation from other studies [14,15]: the ability to detect DEGs is strongly associated with the length of gene as longer genes potentially have more aggregated tag counts.
The denominator of the WaldLog test statistic is: , which is the estimated standard error of the fold change (log scale). If we change the square root to: , then when k = 2, it is the usual WaldLog test; when k→∞, it becomes the fold change criterion as the denominator equals to 1. When the count data has an overdispersed Poisson distribution, the above estimator underestimates the standard error and we will have inflated type I error rate. Under this situation, we need to derive a new estimator for the standard error; or a Bayesian approach assuming the standard error follows a specific distribution can be built. It should be noticed that for the Bayesian approaches, choosing appropriate prior distributions are critical and may have huge impacts on the results. We may also choose models other than the Poisson model. For example, if there are biological replicates, the count data may have overdispersed variances, other models such as overdispersed Poisson distribution or negative binomial distribution can be used [8,9].
Conclusions
When detecting the difference of two Poisson means, if the Poisson means are large, all the statistical tests mentioned above have very similar power. However, their powers may differ a lot when the Poisson means are small. When the count data in RNAseq can be reasonably modelled as Poisson distribution, it is desirable to choose a statistical test which outperforms others for low expression genes. Through simulation study and real RNAseq data analysis, we have shown that the WaldLog test is more attractive than other methods and should be used to identify DEGs.
Methods
Likelihood ratio test (LRT)
Suppose there are n_{1 }and n_{2 }technical replicates for condition 1 and condition 2, respectively; denote the count of short reads of a gene of the j^{th }replicate under condition i by x_{ij}, where i = 1,2, j = 1,2, ..., n_{i}. We also assume random variables X_{ij}~Poi(λ_{i}).
The test statistic of LRT is:
Under the null hypothesis that the two Poisson means are the same, the test statistic has an asymptotic Chisquare distribution with degree of freedom (df) equals to 1.
LRT has been used by Marioni et al [3] based on a GLM, where they first normalize the data across lanes by a global scale normalization method so that the total counts on each lane are the same after normalization.
Conditional binomial test
It can be shown that the conditional distribution of the sum of X_{1j }given the total sum of X_{ij }has a binomial distribution [10,16]. More specifically,
Under the null hypothesis, and the exact pvalue can be calculated from the binomial distribution Bin(n_{1}+n_{2}, p_{0}).
Wald test
The Wald test statistic is [11]:
Under the null hypothesis, T_{Wald }has an asymptotic standard normal distribution.
Wald test for logarithm transformed data (WaldLog)
The Wald test can be also applied to the logarithm transformed count data [11]. We call it WaldLog test. The test statistic can be derived by using Delta method and has the following form:
Variance stabilizing transformation test (VST)
The VST statistic [10,13] has the form:
Fisher exact test
Fisher exact test is applied to the following 2 by 2 table (Table 4):
Table 4. The 2 by 2 table for Fisher test.
The exact pvalue is then calculated based on the conditional hypergeometric distribution.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
C designed the algorithm, conducted the study, and drafted the manuscript; JL analyzed the data and assisted in programming; YD coordinated and directed the whole project. HKTN, SN, HLK, JYY and YD participated in the analysis and discussion. All authors read and approved the final manuscript.
Acknowledgements
ZC would like to thank the support from the NIH grant (UL1 RR024148), awarded to the University of Texas Health Science Center at Houston.
This article has been published as part of BMC Systems Biology Volume 5 Supplement 3, 2011: BIOCOMP 2010  The 2010 International Conference on Bioinformatics & Computational Biology: Systems Biology. The full contents of the supplement are available online at http://www.biomedcentral.com/17520509/5?issue=S3.
References

Bullard JH, Purdom E, Hansen KD, Dudoit S: Evaluation of statistical methods for normalization and differential expression in mRNASeq experiments.
BMC Bioinformatics 2010, 11:94. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Ingolia NT, Ghaemmaghami S, Newman JR, Weissman JS: Genomewide analysis in vivo of translation with nucleotide resolution using ribosome profiling.
Science 2009, 324(5924):218223. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Marioni JC, Mason CE, Mane SM, Stephens M, Gilad Y: RNAseq: an assessment of technical reproducibility and comparison with gene expression arrays.
Genome Res 2008, 18(9):15091517. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Wang Z, Gerstein M, Snyder M: RNASeq: a revolutionary tool for transcriptomics.
Nat Rev Genet 2009, 10(1):5763. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Robinson MD, Oshlack A: A scaling normalization method for differential expression analysis of RNAseq data.
Genome Biol 2010, 11(3):R25. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Bloom JS, Khan Z, Kruglyak L, Singh M, Caudy AA: Measuring differential gene expression by short read sequencing: quantitative comparison to 2channel gene expression microarrays.
BMC Genomics 2009, 10:221. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Wang L, Feng Z, Wang X, Zhang X: DEGseq: an R package for identifying differentially expressed genes from RNAseq data.
Bioinformatics 2010, 26(1):136138. PubMed Abstract  Publisher Full Text

Robinson MD, Smyth GK: Moderated statistical tests for assessing differences in tag abundance.
Bioinformatics 2007, 23(21):28812887. PubMed Abstract  Publisher Full Text

Wu Z, Jenkins BD, Rynearson TA, Dyhrman ST, Saito MA, Mercier M, Whitney LP: Empirical bayes analysis of sequencingbased transcriptional profiling without replicates.
BMC Bioinformatics 2010, 11:564. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Gu K, Ng HK, Tang ML, Schucany WR: Testing the ratio of two poisson rates.
Biom J 2008, 50(2):283298. PubMed Abstract  Publisher Full Text

Ng HK, Tang ML: Testing the equality of two Poisson means using the rate ratio.
Stat Med 2005, 24(6):955965. PubMed Abstract  Publisher Full Text

Ng HKT, Gu K, Tang ML: A comparative study of tests for the difference of two Poisson means.
Computational Statistics and Data Analysis 2007, 51:30853099. Publisher Full Text

Huffman MD: An improved approximate twosample Poisson test.
Applied Statistics 1984, 33:224226. Publisher Full Text

Oshlack A, Wakefield MJ: Transcript length bias in RNAseq data confounds systems biology.
Biol Direct 2009, 4:14. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Young MD, Wakefield MJ, Smyth GK, Oshlack A: Gene ontology analysis for RNAseq: accounting for selection bias.
Genome Biol 2010, 11(2):R14. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Przyborowski J, Wilenski H: Homogeneity of results in testing samples from Poisson series.