Is Data filtering step mandatory?

[smk_77]

Hi All,

I have been trying to use Tophat and also RSEM on illumina data. I always filter the data first when I use Tophat and I have always wondered whether it does its own filteration step or it would be better to filter the data yourself and then trim it using FASTX toolkit etc. Using this approach and then using cuffdiff the overlap with qrtpcr has not been good. So I started using RSEM and I thought I made a mistake by not filtering the reads but then I found that RSEM does neglects the reads by itself as is evident from its output.

# reads processed: 54476831
# reads with at least one reported alignment: 45346199 (83.24%)
# reads that failed to align: 9054059 (16.62%)
# reads with alignments suppressed due to -m: 76573 (0.14%)

Thus I was wondering whether my analysis with RSEM would be correct or wrong. As ideally such tools should automatically filter the reads if they are not bad.

[arvid]

It's not clear to me what you mean by filtering; do you mean trimming off low-quality bases and dropping short/N-containing reads, or something else?

You might want to know what those 16 % non-mapped reads are - if they contain adapter sequences it is lost data that could possibly be recovered.
It is not really necessary to trim off low quality 3' ends with RSEM, since it allows for many mismatches outside Bowtie's seed region (unless they've changed that recently). However, I'd clip off adapter sequences (if any suspected when examining the data with e.g. FastQC) and remove rRNA reads to avoid total read count biases (rRNA tend to be abundant and sometimes at quite different amounts in different samples).