Cufflinks/cuffdiff for matched pair samples and other questions

[azalea14]

Hi all, I am quite new to RNA-seq and bioinformatics in general, so I apologize if these questions are poorly phrased or rather naïve.

(1) I have paired RNA-seq data (normal vs. cancer) for several patients and have begun DEG analysis through the tophat-->cufflinks-->cuffdiff pipeline. I’m wondering if this is even a viable pathway, as a thread regarding an earlier version of cuffdiff expressed some doubts about its suitability for paired designs: http://seqanswers.com/forums/showthread.php?t=7108. Does the latest version of cufflinks/cuffdiff address these problems, or should I consider a different approach?

(2) This is probably a dumb question, but what is wrong with just importing cufflinks’ FPKMs for all my samples into Excel and running a paired t-test on them to determine differential expression? As you can tell, I am not well-versed in the statistics of this...

(3) Assuming I can use cuffdiff, I’ve been encountering issues with the FPKMs it produces, which are all much, much larger than those given by cufflinks for the samples I’ve tried. I know inconsistencies like these have been caused by different default settings between cufflinks/cuffdiff in the past, but several threads have mentioned that the latest version (v2.1.1) should have this fixed. Am I doing something wrong? Below is my code and sample output.

For cufflinks:

cufflinks –G reference.gtf patient_normal.bam

In genes.fpkm_tracking:

gene_id   locus         FPKM     FPKM_conf_lo   FPKM_conf_hi     FPKM_status  
gene X chr1:123596-123889 0.04095203     0.0225433     0.0751442  OK

For cuffdiff:

cuffdiff reference.gtf patient_normal.bam patient_tumor.bam

In genes_exp.diff (value_1 should refer to FPKM for normal sample):

 
gene_id  locus         sample_1  sample_2 status  value_1  value_2  log2(fold_change)  test_stat  p_value  q_value  significant                        	
gene X  chr1:123596-123889  q1     q2     NOTEST  16.7791  17.0898  0.0264722                1        1       no

Thank you for all your help!

[dpryan]

(1) Don't use cuffdiff for that sort of analysis, it only handles pairwise and time-series designs. Look into DESeq2, edgeR, or limma, all of which can handle your design.

(2) Are FPKM values normally distributed (have just looked at a few, I wouldn't assume that)? Also, unless you have quite a few samples, you'll benefit from sharing information between genes (this is done in the aforementioned packages).