Denovo RNA-Seq assembly using Velvet/Oases

[AdrianP]

Hello dear colleagues,

I feel like I should know the answer to what I am about to ask but I could not find a paper with a clear cut answer to what protocol I should apply. I am getting down to it:

I have illumina pair-end reads from RNA-Seq of a fish that is very distant from anything sequenced. I have about 100 mil reads in both directions.

I used Velvet Optimiser to perform denovo assemblies on has lengths 67 77 83 and 93 in order to look at the results and chose the best assembly. However, this is what I encountered (velvet):
Hash 67 has about 250k contigs
Hash 77 has about 125k contigs
Hash 83 has about 78k contigs
Hash 95 has about 15k contigs

So not that big of a suprise, lower hashes more contigs, higher hashes are stricter because of their length, so less contigs.

I polished these using Oasis, and the numbers did not change a lot.
250k 149k 93k 16k
n50s: 867 897 889 827

I tried the -merge function suggested in the Oasis manual, and got 324k contigs and a bit higher n50 about 1000.

Here comes the question, now what? which one is the best assembly? they can't all be as good. Which one should I annotate?

Any help? I am very stuck.

[AdrianP]

bump. if anyone has any input i would love to hear it.

[CPCantalapiedra]

just asking, have you tried other software, like trinity to compare and get more clues for data interpretation?
Maybe you could try too to compare some of the contigs with very highly conserved proteins among vertebrates?

[vbiaudet]

You could consider each kmer, I used velvet and oases with each kmer in descending kmer:
I began with kmer=95
--> take the contigs and us it (-long parameter) in the next velvet/oases with kmer=83
--> continue this until the small kmer and you obtain one list of contigs

After you have to use an other assembly of contigs as TGICL or iAssembler

I applied this method to Arabidopsis transcripts and find 90% of contigs that are good compare to gene model...

[AdrianP]

how far down do I go?

Can anyone else explain to me what confidence is? I understand that it is when there are alternatively spliced transcripts, but what does a very low confidence mean?

[2seq]

bump...would like to hear more about this. I also like vbiaudet 's idea very much for getting one list of contigs. However, how does this message differ from the oasis merge function?