to reuse or not reuse spade corrected reads - that is the question...

[evezeyl]

As corrected reads are output from the bayesian hammer error correction from spades -> I would reuse those instead of using the original trimmed reads.
This error correction appears sensible as it uses clustering of K-mers - so IF kmer is very rare its is assumed that it is an error unless the phred quality is really good. it appears to be conservative and a good method for data with good coverage. Errors are corrected only when phred is not good and only few representant in cluster. So, quiet conservative.
we are likely to reintroduce error if we used trimmed reads.
Aslo it appears to be more consistent to reuse those reads, as we actually generated the assembly with those corrected reads. (But true enought that: unless the filtering quality is set very low in trimmomatic, it should not make a huge difference afterwards). will make some tests...

[evezeyl]

I agree with myself now also (still will make the test) - but I think we should use the corrected reads for everything after Spades because:

1. it seems more consistent scientifically (would reintroduce errors while trying to improve assembly)
2. because of post-analyses also would need to be done on a consistent dataset: if if I try to detect SNPs with mapping reads to a reference with the original reads this would then not be equal to using my assemblies and detect SNPs with the same reference (meaning those methods would not be directly comparable because I would have worked on slightly different data sets). + feks. i would likely detect less variants while using original reads (in regard to my corrected assembly) because of rare inconsistencies eliminating those poditions.
   As several software do their error correction, I would actually doing this error correction only once (with the spades bayeshammer - apparently one of the best method so far) - and deactivate error correction in other software used after (recorrecting the reads with a different method does not seem appropriate either, because again this might lead to work on a slightly different dataset). Other solution of course is not to correct at all and be more stringent in the original filtering (but then we might loose more coverage) .
   Well letting you think about those aspects. No more issues :)