Merge from multiple libraries + tophat + cufflinks

[vinay052003]

Hi,
I have two paired-end sequencing RNA-Seq libraries. Both of them are from the same sample but prepared using different insert lengths. I've run tophat on both of them separately. Now I want to run Cufflinks. I was wondering if it is a good idea to merge tophat accepted_hits.bam files for different insert lengths and run Cuffllinks on the merge file?
My goal is to detect splice variants for the whole sample.

Thanks

[sdriscoll]

I don't think it will matter in the end. If you are going to use Cufflinks then with two separate alignments you'll have to use cuffmerge to merge the separate transcript annotations into a single consensus annotation. At that point the information from both runs of the sample will be merged. In fact everything will probably go smoother if you just run them through cufflinks separately and follow that with cuffmerge.

[vinay052003]

That's a good idea. I thought of that as well. Actually in the end I also need the expression of each transcript. And after merging assemblies, in order to estimate the expression, I'll need to use two bam files anyways. That's why I was wondering weather to mege bam files before running the Cufflinks or to do it later.
Any thoughts?
Thanks.

[sdriscoll]

Cufflinks needs to know your insert size and standard deviation. In fact it will figure it out from your alignments. Since that information is apparently necessary for it to do it's work then you should probably keep those alignments separate through cufflinks. I get you, though, you want one set of expressions not two. Right?

[vinay052003]

Yes.. Is there any good way?

[sdriscoll]

I'm not sure if you need anything much more exotic than a mean of the two separate expressions.

*edit* scratch that. a mean isn't quite right. calculating expressions is pretty easy when you have the raw components and when you have those you can easily combine expressions but when you only have two separate expressions then i'm not sure how to do it.

for example, say we have some gene X of length 3400 bases.

sample A has 12000 reads of aligned to the gene, 34,123,123 reads aligned
sample B has 5161 reads aligned to the gene, 25,321,321 reads aligned total

sample A RPKM = (reads*1e9)/(reads_aligned * length) = (12000*1e9)/(34,123,123*3400) = 103.43
sample B RPKM = (5161*1e9)/(25,321,321 * 3400) = 59.95

mean of those expressions = (103.43+59.95)/2 = 81.69

RPKM of merged samples = ((12000+5161)*1e9)/((34,123,123+25,321,321)*3400) = 84.91

the weighted mean works but you need to know the # of millions of mapped reads to use as the weights:

weighted mean of samples = (103.43*34,123,123 + 59.95*25,321,321)/(34,123,123+25,321,321) = 84.91

i don't think cufflinks reports what the normalization factor is it uses for FPKM calculation so i'm not sure you'll find an easy way to merge them. I guess you have no choice but to run them combined as a single BAM.