2QC or !2QC and whether to check for adapter pollution?

[LeonDK]

Hi guys,

In the case of mapping Illumina reads to a reference genome, I was discussing with various people whether to perform QC on raw fastq files or not, I get the impression that there is a 'for' and a 'against' stand point amongst NGS scientists... As I understand it, there is basically 2 point of views:

1) NO, do not QC! The aligner will take care of the bad quality reads and 3' quality deterioration and adapter contamination by soft clipping. The aligner STAR was mentioned in this context

2) YES, do QC! There is no reason to 'confuse' the aligner by introducing reads and bases and adapters, which are known to be of bad quality/technical artefact. Basically why keep noise, if it is easily identified and removed.

I'm interested in hearing your take on this matter?

Cheers,
Leon

[dpryan]

In the case of RNAseq it probably doesn't matter too terribly much. The general best practice is to trim adapters (you can be pretty liberal at leaving contamination in) and then optionally perform minimal quality trimming (not removing anything above phred=5). This is assuming that you're using something like STAR that uses local alignment.

For other *seq experiments the situation can be completely different. With bisulfite sequencing, for example, you'd just decrease the quality of your data by not aggressively trimming adapters (though even there I don't think aggressive quality trimming is particularly useful). It's all a question of how much error you can tolerate in downstream analyses.

[Brian Bushnell]

Whether an aligner is local or global, adapter sequence in the data is never going to improve results, and always poses a risk of spurious placement and bad alignments. If you can remove contaminant sequence with high precision, you should always do so prior to any analysis. Even basic questions like "What percent of my DNA came from this organism?" can't be answered usefully if, say, 50% of your data is adapter-dimers. Which you would never know without QC.

Quality-trimming is another matter, and as Devon said, it depends on how much error can be tolerated downstream. Mapping, for example, can be more accurate when including low-quality bases; if you have a 120bp repeat element, a 150bp read with 50 low-quality bases may map uniquely while a 100bp trimmed read will not. But assemblers may be less tolerant.