| DESeq2 [32] |
Empirical shrinkage estimation of dispersions and logarithmic fold-changes. Z-test is used for DE analysis. Both outliers of dispersions and logarithmic fold-changes are treated. |
DESeq2 exhibited steady and good performances regardless of outliers, sample size, proportion of DE genes, dispersions, and mean counts. |
| edgeR.rb [28] |
Observations that deviate strongly from the model fit are given lower weights. These observation weights affect both the regression and dispersion estimates. Used when data include outlier counts. |
edgeR.rb yielded more DE genes and more false positives compared to other edgeR methods. In presence of outliers and large number of samples (≥10), it exhibited outperforming AUCs. |
| edgeR.ql [29] |
While edgeR exact test assumes the estimated dispersion is true, quasi-likelihood estimation accounts for the uncertainty of the dispersion estimates. This approach improves type I error control. |
edgeR.ql showed better AUC, control of true FDR, and FPCs compared with edgeR methods, but exhibited relatively low power. |
| voom.qn /voom.tmm [34] |
Read counts were quantile normalized (voom.qn) or normalized with TMM method (voom.tmm), and then were transformed using voom. A moderated t-test is used for DE analysis. |
voom.tmm performed better than voom.qn except outlying sample cases. They exhibited overall good performance for most cases, but their powers were relatively low. AUC of voom.qn was noticeably decreased compared to other voom methods as the proportion of DE genes increased. |
| voom.sw [36] |
Observations from highly variable samples are down-weighted for more accurate estimation of regression coefficients. Used when some samples have amplified dispersions |
voom.sw performed like voom.tmm rather than voom.qn and showed overall good performance. When samples with amplified dispersions were included, voom.sw outperformed other methods. |
| ROTS [19] |
voom transformation with TMM normalization and bootstrap are used for generalized t-statistic that maximally reproduce preselected top k% genes. k = 25 was used for our tests. |
ROTS exhibited good AUC and false positive control. ROTS applied to raw count data showed slightly lowered power, but it outperformed other methods when DE genes were unbalanced and a large number of DE genes were included. |
