dealing with 5' end bias in RNASEQ

[PFS]

Hello,
I read about random hexamer priming bias in Illumina RNASEQ (see Hansen et al., NAR 2010, vol 38 no 12). Specifically, "There is a strong distinctive pattern in the nucleotide frequencies of the first 13 positions at the 50-end of mapped RNA-Seq reads".

I assume that is the reason why my FASTQC report does not show parallel lines in the first 13 positions for GC content.

How do you deal with it? Will trimming work?

Thanks!

[PFS]

I read more carefully the paper, and I understand that trimming the 5' end does not work.

I am curious to know whether this bias is of any real concern for differential expression analysis. If so, do you use Genominator?

[sdriscoll]

Personally I find the bias confuses things. After you run your reads through an aligner and process the abundances of reads aligning to genes you'll find 10's of thousands of genes with some expression. Figuring out which genes are "actually expressed" can be a little confusing. It doesn't help when the coverages are skewed due to a 5' bias. I've analyzed a run that had such a strong bias even genes with expressions in the 1000 FPKM range were not fully covered by reads. Does that mean the gene wasn't there? Probably not - but it certainly confuses the issue. I've also heard of the trimming idea - didn't work for me. We just took the expressions we got and went with it.

[pmiguel]

Does anyone see this with the Roche Rapid cDNA prep and 454 sequence? I would think so, because it also uses hexamers for cDNA synthesis/2nd strand synthesis. But if not, it might be that a minor change in the cDNA synthesis protocol could remove the bias.

How about using random 9-mers?

Or "balanced" 6-mers? (Where all 4096 possible six base sequences are synthesized separately, quantitated and pooled to ensure equi-molar amounts of each 6 base sequence.)

--
Phillip

[Simon Anders]

I am curious to know whether this bias is of any real concern for differential expression analysis.

No, not really. The bias will affect estimates of absolute expression, but once you calculate a fold change for a gene by comparing several samples, it should cancel out.

This holds if the patterns are the same in all samples. If they are not, you might get better results when adjusting for it. This is at least what Hansen et al. claim in their follow-up paper, a preprint of which you can find here: http://www.bepress.com/jhubiostat/paper227/