Hope this doesn't raise more questions than answering yours, and this is just my personal experience. I must say, Blat gave me quite a good impression in the past. It has support for large inserts, controlled via maxIntronSize and fine parameters. I am not aware of another tool that gives you control over that so easily. And it also seems to be able to align more reads to my reference than many of the dedicated short-read tools, and no need to build an store an index. What's not so good: e.g. no SAM/CIGAR output format, not multithreaded, the licence, etc. And it can be beaten in sensitivity and speed.

I made a little comparison, took a random sample (maybe some 10.000 for speed) of my reads in fastq and fasta and simply try how many reads can be mapped with different aligners, caveat: didn't care for alignment quality (should check for validity though), and no Introns, and maybe it's totally screwed, here is my list, your mileage may vary:

1. BFAST (~85%) (BLAT-like fast accurate search tool)
2. LastZ (~80%)
3. Novoalign (~80%)
4. Blat (~74%)
5. Bowtie (~60%)
6. bwa (~54%)

I recommend: Bowtie&Tophat: because it is developed specifically for that, LastZ: has hell of a lot parameters, SAM output etc., BFAST: because it gave me most alignments, was the slowest though.

You have to find out, how to set parameters to allow for large inserts/introns though, try to play with the gap costs. Would be very nice if you post your experiences!

Just heard of this SpliceMap tool from stanford, not mentioned in any of the above lists: http://www.stanford.edu/group/wonglab/SpliceMap/index.html

Some discussion of the tool: http://seqanswers.com/forums/showthread.php?t=5507

I've had good luck with tophat (and paper) and have recently been trying out gsnap (and paper).

Both can find novel splice sites and/or take a list of known splice sites.

gsnap seems to find more splice sites and is more tolerant of indels especially if you have longer reads.

We've created a pipeline using bowtie and blat which maps against the genome and transcriptome and merges the results. I'll be presenting this pipeline at the MGED conference in Boston in a poster. Using simulated data we are achieving 10% more accurate alignments than any of the other methods, we compared to TopHat, GSNAP, SpliceMap, BWA, NOVOALIGN, SOAP. TopHat has a good false positive rate for junctions, by basically just calling the really easy stuff, so it has a very high false negative rate, it was something like 1% false positive and 50% false negative, compared to 5% false positive and 10% false negative. We were also going to use BFAST but was told by the author Nils Homer the following: "BFAST is meant for whole genome resequencing, so it will not be able to find splice forms unless those splice forms are found in the reference. Software like cufflinks or tophat would be more appropriate." If anybody wants to use our pipeline email me, ggrant@pcbi.upenn.edu

First thing to look is the quality of your RNA-Seq reads. Check the thread:

"What is the best pipeline for human whole exome sequencing?"

You may also think about using some hybrid approach:

• map everything in an unspliced mode with last (gives more hits than blat for the same set, but some are unrealistic -> too many indels)
• check few different cutoffs, for what is reliable mapping (3 indels?)
• map all unmapped & unreliable hits with few spliced mappers, compare & combine the results. Incomplete list

I think there is a lot of potential in this area and worth trying out different approaches. We have been trying to gain more traction with RNASeq data and Novoalign (we're the developers) and it seems there is room to combine our mapping strategy with an approach like Tophat's or SplitSeek.

I will be happy to share more once I start seeing some promising results. From what I can already see there are going to be many cases where you need to consider gapped alignments in short reads.