Extracting reads with highest identity to my genes of interest from illumina raw data

[ataheri]

I was wondering if there is a way to pull out the hits from illumina raw data according to a set of genes I am interested in rather than doing assembly and alignment for all the raw data.
I have 66 paired-end RNA-seq samples and would like to only look at 80 genes.

[Richard Finney]

Yes, it's possible. In fact it's fairly trivial. Just think of the "genes" as "chromsomes", build and index and align. There are problems:
1) "non-canonical" splice forms
2) alternative splicing
3) genes in a gene family (or that are not unique in terms of sequence "mapability").
4) what if it's really an immature rna and maps to the genome?

It turns into a real headache. Plus everyone munges their bam files and tools are not inter-operable.

Nobody's really figured out RNA-seq with short reads ... completely figured out that is. There are some "good guessers" out there.

[ataheri]

So, in this case which tools do you think works better than the others?

[swbarnes2]

Unless you intend to do a hell of a lot of grep, you have to align.

You can align to a small genome consisting of only your genes, and maybe set the stringency high, so that reads from other sites of the genome aren't forced to ailgn badly to the limited reference you've given them.

[Richard Finney]

So, in this case which tools do you think works better than the others?

Caveat: my opinions.

Order of goodness:
1) Roll your own: align to 1) genome and 2) a build your own reference gene model , align using (in order of goodness) either novolign, bwa, bowtie ... then write your own "model vs. genome pick the best alignment" tool and fine tune it to your liking. VERY DIFFICULT.
2) Rum
3) Mapsplice
4) GSNAP

I have not checked out bowtie/cufflinks nor scripture.

[Simon Anders]

What is wrong with simply aligning to the whole genome, as it is usually done with RNA-Seq data, and then looking at your genes? Aligning is not such a computationally expensive process.

Aligning to a reduced reference will cause you to get spurious mappings of reads that actually stem from similar regions missing in your reference, and it won't be much quicker, anyway. (Alignment speed should depend only logarithmically on reference size.)