Trade off between read length and number of reads

[ymc]

For differential expression analysis, more reads means more accurate expression level estimate and longer read length means more accurate mapping to the genome and the transcriptome.

Obviously, if money is no object, increasing both parameters should be a no-brainer. But if we have to choose between doubling number of reads or doubling the read length, which one will be the better choice for differential expression analysis? Why?

[Simon Anders]

Depth over length.

Even with 30 bp read length, most reads map uniquely, because there is not that much repetition in coding regions. Longer reads give only a marginal improvement. If you want to discern isoforms, read length might become more valuable, though (even though good bioinformatics support to make efficient use of this extra information is still not mature, in my opinion).

Even better than depth are extra replicates. Remember that your power to detect differences depend on the number of reads not per sample but per condition.

[Jeremy]

If you already have a reference genome available then you really can't go past Illumina, the read length that you get from them now (2x100 bp PE) is fine for aligning to a genome and getting good results.
I have used 454 without a reference genome available and the amount of data that you get, while perhaps more easily assembled, is not very good for coverage.

[ymc]

Is around 123M 2x100 reads roughly the standard for human RNA-Seq? I found that Stanford's iPOP project sequences 123M 2x100 reads in each of its ten time points.

http://www.stanford.edu/~gmias/iPOP/data.html