Comparing Fpkm values from samples processed in same or different runs using tophat/cufflinks

Question

[trakhtenberg]

I have 12 pair-end samples, I would like to compare just Fpkm values using tophat/cufflinks. Would the Fpkm values be the same (for annotated and/or novel transcripts) regardless whether or not I process all 12 samples in the same run (i.e., put 12 group folders into the main folder in which I run the program) or if I run 2 samples at a time and then just compile all 12 in one table?

Could I process and compare in the same run pair-end samples sequenced 50bp or 100bp from each end? Would the platform affect the Fpkm values or I can assume normalization is appropriate and I can rely on differential expression results?

Answer 1

[Devon Ryan]

They won't be the same if you have novel transcripts and then use multiple runs. There's a step where one merges (via cuffmerge) the annotations produced by cufflinks for each sample. The FPKMs are dependent upon that merged annotation, so at least if you perform the merge step with different groups of samples then you're pretty much guaranteed to see differences. Further, I think cuffdiff defaults to using a DESeq-like library size normalization rather than the old and less reliable "divide by the number of aligned reads" method. That can also produce (generally pretty small) differences in FPKM estimation.

Regarding using a mix of read lengths, yeah, you can do that (you can also mix paired-end and single-end). Keep in mind that mappability will change a bit, so that'll produce a bias. If you have multiple platforms, then that'll be a confounder. I should note that cuffdiff is really only designed to handle very simple experimental designs (i.e., two group comparisons or a time series). If your experiment doesn't fit into one of the defaults, you'd be better off using a different tool (just use one of the count-based methods, like DESeq2/edgeR/etc.).

[trakhtenberg]

thank you!!!