Differential Expression across libraries/platforms

[mikecz]

I've been searching for a while now and I can't seem to find any kind of discussion on this topic.

Is there any consensus on the most appropriate way (if any) to calculate differential expression of transcripts with samples derived from different library preparation methods or from different sequencing platforms?

I have a few Illumina Hiseq 2000 100bp PE RNA-seq samples that can be considered biological replicates, but not quite technical replicates since some are from Truseq stranded libraries with others from Truseq unstranded preps. The problem is more complex yet since the unstranded samples have ~150 million read pairs and the stranded samples only 25 million.

So far I've used the stranded reads for de novo transcriptome assembly, then used both the stranded and unstranded reads for mapping and quantification. Treating them in the same way by ignoring strand information with Bowtie and allowing for unpaired mapping when concordant mapping isn't possible gives me pretty horrible expression correlations after RSEM.

Is there a normalization method that can compensate for the different biases in library preparation?

[dpryan]

Just treat all of the samples as you would normally and treat the technology type/library-prep-kit as a batch variable (or two) in your model design.

Of course, if whatever you're actually interested in measuring is partitioned across these then you're probably SOL.