Different RPKM values in same dataset using Cufflinks or Cuffdiff (v1.3.0)

[hlwright]

I have generated RPKM values for the same dataset using Cufflinks (single input BAM file) and using Cuffdiff where my datasert is part of a larger set of samples (multiple input BAM files). The commands are as follows:-

cufflinks -b /path/to/BowtieIndex/genome.fa -p 8 --max-bundle-frags 100000000 -G /path/to/genes.gtf sample1.bam 

cuffdiff -b /path/to/BowtieIndex/genome.fa -p 8 --max-bundle-frags 100000000 /path/to/genes.gtf sample1.bam sample2.bam sample3.bam sample4.bam sample5.bam sample6.bam 

I have noticed that when I compare RPKM values for individual genes (genes.fpkm_tracking file), the RPKM values generated by the two programs are different even though the input BAM file is the same, for example:-

Gene 1
Cufflinks RPKM 1138.91 (confidence intervals 1130.66,1147.16)
Cuffdiff RPKM 1572.47 (confidence intervals 637.99,2506.94)

Gene 2
Cufflinks RPKM 58.15 (confidence intervals 56.40,59.89)
Cuffdiff RPKM 78.99 (confidence intervals 39.83,118.15)

Can anyone explain this please?

[sdriscoll]

This is by design. I think it would be wiser and more science like to use the following type of pipeline:

Align reads once for expression analysis.
Quantify gene expression once as either read counts, cufflinks estimates or try RSEM
Use a DE tool like DESeq, edgeR or the more recent EBSeq which seems to have improved on previous methods a bit.

It doesn't make sense to get multiple expression estimates from the same data even though it makes sense from the computation logic side of things.

I think RSEM might be my new favorite except that it doesn't run well on my Mac system. If you have a strong Linux system then it should be good. If you run RSEM multiple times on the same data you will see some variation in its isoform level assignment of expression however that's a result of variable/random behavior of the aligner. The gene level estimates are more stable. This only exposes the fact that we have all probably been dealing with this extra uncertainty in expression values all along. Their pipeline requires bowtie and they run it in a specific way for good quantification estimates. So if you want to try the RSEM pipeline you only need to do that and you can skip the initial alignments because RSEM provides them for you. You would run it once per sample and then merge the data for DE analysis. They recommend EBSeq, in fact they package it with their software.

[sdriscoll]

today I discovered eXpress (http://bio.math.berkeley.edu/eXpress/overview.html) which uses the same basic algorithm as RSEM but is much faster and it produces more verbose output. so far I like it and I've seen that it's expression estimates correlate very highly (r > 0.8) with 'true' expressions from synthetic data analysis. someone shared a slideshow with me outlining an evaluation of current possible pipelines using the BEERS pipeline (http://www.cbil.upenn.edu/BEERS/). cufflinks wasn't even on the map with count estimates correlating 0 < r < 0.2 - or in other words the expression estimates looked like random noise compared to the true values.

[hlwright]

Thanks sdriscoll - I have been getting more and more frustrated with cufflinks / cuffdiff so will explore these other options.

Really appreciate you replying to my post.

Helen