I am going to do RNAseq on multiple conditions with 3 replicates for each, to de novo assembly the transcriptome and identify some low expressed factors (<1% in transcriptome) in different conditions. Quantitative comparison is required, and detection of isoform changes even SNPs are preferred.
Hiseq with 76 or 100bp PE reads is obviously the first choice. Since we have 12 conditions, plus 3 replicates, 36 samples will be sequenced in total. If a depth of 100-200 M reads per sample is recommended for sensitive detection, it means we need at least 9 lanes for all my samples. That will be a huge expense.
I am wondering whether it is possible to reduce depth to 40-50 M reads per samples, considering I have 3 replicates. Or what kind of design (depth, replicate arrangement in lanes ...) can be an alternative plan with least compromise on my purpose but lowest price?
BTW, the genome size is estimated as 1300 Mb.
I think you'll probably be fine with 45-50 million reads. Don't quote me on this (going from memory of an Ilumina white paper I read awhile back), but 1.5 million reads gives equivalent sensitivity to a microarray, 10-15 million reads is low-depth, 30 million reads is medium sensitivity, and anything over 45 million reads is considered "deep sequencing." You can do a quantitative comparison with almost any of those, and calculating differential splicing is a piece of cake (look up the Cufflinks suite). The only reason you'd want to go above 45-50 million reads is if you were REALLY interested in SNPs. I'd definitely go with 100bp reads for de-novo assembly.
Are you doing the bioinformatics yourself? Because if not, you should learn how to. Learning how to do RNA-Seq bioinformatics is fairly easy and it cut the price of my project by more than a third. In your case, you could probably save even more money, as all I was doing was calculating differential gene expression. By figuring out how to do the bioinformatics yourself, you can sequence more samples/get more read depth for the same price (because you don't have to pay for the bioinformatics).
So if I were you, I'd aim for a read depth of 40-60 million reads, get 100bp paired end reads, and do the bioinformatics yourself if money is an issue (I cannot overstate how much money you will save).
for de novo assembly of transcriptome, you need 50-100 million pair-end reads.
after you have a good reference transcriptome, you can replicate your different treatment with 15- 20 million reads for differential expression. see here http://angus.readthedocs.org/en/2014/_static/ngs2014-trimming.pdf