New differential testing of cuffdiff/cufflinks since 1.3.0

[sudders]

I am also having the same problem in both splicing.diff and promotor.diff.

I have 5 samples in each of two groups, but only 25M reads per sample (100bp paired end).

Promoters

Cufflinks 1.0.3

OK 11,309
NOTEST 8,637
FAIL 648

with 7,902 significant

Cuffdiff 1.3

OK 1,383
NOTEST 10,243
FAIL 8,948

2 significant

Splicing

Cufflinks 1.0.3

OK 13,996
NOTEST 51,240
FAIL 2,933

8,645 genes with sig. differential splicing

Cufflinks 1.3

OK 1,017
NOTEST 28,032
FAIL 37,736
LOWDATA 1,401

6 genes with sig. differential splicing.

[DerSeb]

Hi all,

I have been trying out the new 1.3 version of the cufflinks package to see if the testing for differential splicing/expression etc is improved.

Now I expected a more stringent analysis, as announced, but I get mostly failed tests.

I used the same reads.bam and annotation and the same parameters (5 biological replicates and 2 groups, 2x 100 bp PE reads, ~ 140mil per sample).

This is my result with cuffdiff v1.0.3 for diff. promoter usage:

191 FAIL
13713 NOTEST
4745 OK

with 1403 Genes found with sign. diff. promoter usage.

Now this is the cuffdiff v 1.3 result:

5719 FAIL
35 LOWDATA
12301 NOTEST
594 OK

with 20 Genes found with sign. diff. promoter usage.

It seems that the number of false positives is really reduced. The new version find only ~ 3.5% of genes tested to be differential. However, so many fail to be tested in the first place that the result is not really usable.

Did anyone experience something similar? Is this failing of testing maybe related to my large dataset or the replicates?

All the best,
Sebastian