| Fold change |
– Calculates the ratio of a gene's expression between sample and control |
– Works with small sample size |
limma64,65
|
| – Genes are classed as differentially expressed according to a selected threshold (usually an absolute log-fold change value greater than 0.5 to 2) |
– Easy to interpret |
WAD66
|
| – Usually used in conjunction with a non-parametric/linear/Bayesian significant test |
– Does not take into account the sample variance |
|
|
|
– Different ways to calculate depending on the use of averages, medians, etc.
|
|
|
|
| Non-parametric tests |
– Rank-product method, Mann whitney U tests for comparing two categories |
– Capable to compare different platforms’ results |
RankProd68,69
|
| – Kruskal–Wallis test for multiple categories |
– RankProd is best method for meta analysis67
|
| – Compares the ranks of the genes according to their expression |
|
|
|
| Linear methods (t-test, ANOVA) |
– Compare the mean value of expression per gene in samples |
– Add statistical significance, but uses the boundary condition the gene expression values of conditions are normally distributed |
Cuffdiff 270
|
| – The null hypothesis is that the means are equal – t-test is for two category comparison |
– Commonly used with fold change |
limma – after a Bayes procedure |
| – ANOVA is for multiple categories |
|
|
|
|
| Bayesian methods |
– Use the data to predict the probabilities of differential expression |
– Have relatively high computation time dependence |
limma |
| – Use the standard deviation to alter the test statistics or tests directly |
– Makes more appropriate results then a t-test |
baySeq71
|
|
|
| Counting method |
– Uses the real count of the expressions for comparison with a negative binomial test |
– Requires exact number of mRNA copies |
DESeq265
|
| edgeR72
|
