Skip to main content

Table 5 Summary of the main observation for selected methods

From: Comparisons of computational methods for differential alternative splicing detection using RNA-seq in plant systems

 

Class

Novel AS

Detection region

Comments

DiffSplice

IR

Any type

ASM

Assembles transcriptome based on graph theory. Does not rely on annotation but does not use annotation either. The goodness of ASM is questionable. Generally low AUC. Performs poorly when detecting SE events.

Cufflinks

IR

Any type

Gene

Assembled transcripts merge with annotation to provide a more confident reference. Is least affected by incomplete annotation. Model is designed for pair-end data. Performs better for medium read depth than both low and high read depth. Performs better when detecting A3SS and A5SS events than other types of AS events. Computationally slow, but allows parallelization.

DEXSeq

CB

Only SE

Exon

Uses a generalized linear NB model. Achieves the highest AUC in many cases using accurate annotation. However, incomplete annotation can impose considerable problems for it. Poor FDR control.

MATS

CB

NS

AS event

Uses a Bayesian model. Solely based on junction reads. Can not detect complex AS events. Annotates splicing events with corresponding event types. Good FDR control in many simulation studies. Performs the best for real data.

rDiff-param

CB

NS

Gene

Conservative with default settings. Good FDR control but low AUC in many cases. Computationally fast.

SplicingCompass

CB

Only SE

Gene

Compares geometry angles of read count vectors. Generally poor FDR control and Medium AUC. Performs well when detecting SE events.

DSGseq

CB

Only SE

Gene

No p-value reported. Generally medium AUC. Performs well when detecting IR events and when using incomplete annotation. Computationally fast.

SeqGSEA

CB

Only SE

Gene

Integrates DE analysis with DS analysis. Generally high AUC. Requires a sample size around 5 to claim significance at a reasonable FDR level, i.e. F D R=0.05. Computation time increases dramatically as permutation times increases.

  1. IR: Isoform resolution models. CB: Count based models. NS: Not Supported. ASM: Alternative Spliced Module.