Diff expression: pooling reads of replicates/treatments for de-novo assembly

[CarlosVM]

I want to make pairwise comparisons of gene expression between tissue samples (same tissue, same species, different individuals) in 5 different treatments (with 2 biological and 2 technical -sequencing- replicates per treatment, using Illumina paired-end reads for de novo assembly). Before the differential expression analysis, I have to assemble a de-novo transcriptome.
Ideally, I'd like to have a good tradeoff between maximum recovery of splice variants and not too many computational chimeras.
Assuming unlimited computational resources, what would be the best strategy for pooling the samples in order to get a common set of transcripts for which to compare expression in different treatments. I thought of pooling all 20 samples for creating a single assembly that would contain the transcripts expressed in every condition and then I could map each sample to this assembly and subsequently compare. How much coverage is too much? (in terms of errors, chimeric sequences). My main concern is on how this will affect to the representation of isoforms from different treatments.
Is it more appropriate to make 5 different assemblies with 4 samples each and then collapse them with CD-HIT or a similar tool?
Thanks,

Carlos

[Jeremy]

Combined assemblies are the way to go, a few programs even give the information about which read formed which contig in the output. Trinity does this, although I am only just starting to play with it now. It looks promising.

[Jan_R]

Unlimited computational resources.. I'm jealous.
When you talk about 20 different samples, how many reads do you have in total? There is an estimation on the Trinity homepage about the amount of memory needed per amount of reads I think.

The coverage depends on the expression of your genes and on the amount of reads you obtained in each sequencing run. I find it hard to suggest parameters for that.

I did a similar aproach on such a problem. First I asembled with Trinity and then did the alignment with bwa. Trinity is in fact able to give you rpkm values for each assembled sequence. Though last time I looked upon their sourceforge page they were still admiting one shouldn't trust these values. Unfortunatly at least parts of sourceforge seems to be down at the moment so I can't look it up.

[Anemone]

@CarlosVM
How far have you done in your analyses? I am planning to analyze similar things (have 4 different conditions, 3 individuals per each condition and 3 samples per individual - different tissues, together 4 x 3 x 3 = 36 samples). I isolated RNA separately, but thinking to maybe pool the different tissues for each individual before sequencing to get less libraries to sequence (but still sequence from different tissues to get better transcriptome profiles). I also have to do de novo assembly and I'mplanning to just assembly from the reads I sequence, plus some available ESTs online.
Then my thinking was like yours, map each sample back to assembly and compare the samples between conditions.

Do you think the pooling of tissues for each individual separately is a good idea or would you just barcode each tissue separately?
Do you maybe have some advices for the analysis?

We are mainly using CLC Genomics Workbench for our analyses, but I used also i.e. Trinity assembler. The main issue I have with it is that Trinity reports alternative transcripts, so that you actually think you have more transcripts than you actually have and I think that might be problem for some follow up analyses..or? Do you have some experience with RSEM, or edgeR, DESeq programmes?? I'm just trying to read up on it.

Thanks in advance for any advices,
Please ask if something was not so clearly explained,
Ilona