Fastqc 'Chip-Seq' Quality Score Reverse Pattern [Quality Increasing At Ends]

Question

[Sukhi Singh]

Hola! I am encountering a strange problem. My fastqc graph is like this, with quality score increasing at the end, but we should observe is a decrease at the end.

The fastq files are generated using CASAVA-1.8.0 , so the format is supposed to be sanger encoded.

My previous graphs from different experiment they show a decrease in the end as opposed to this one.

Why I am observing this pattern (increase in quality scores at the end)?

Thanks for your comments.

[Niek De Klein]

What is your question?

[Sukhi Singh]

Why I am observing this pattern (increase in quality scores at the end)?

[Steve Lianoglou]

Ever since we've switched to running our samples on a HiSeq machine[*], all of our phred distros exhibit this exact same pattern, and I'd have to say: judging by this quality distro plot alone, your data actually looks pretty great.

[*]I'm not sure if it was the switch to the HiSeq, or the upgraded software/chemistry -- maybe GAIIx runs look like this now, too ... I wouldn't know, though.

[Brad Chapman]

The modeling of error rates changed in recent versions of the Illumina software. See this question for more details: http://biostar.stackexchange.com/questions/12150/rna-seq-difference-in-read-quality-pattern-between-illumina-ga-and-hiseq-2000/12179#12179

[Sukhi Singh]

@Steve @Brad Thanks for your comments, I think the graph is fine, its just the change in error model by Illumina

Answer 1

[Istvan Albert]

I wouldn't read too much into it. As you can see your error bars are actually increasing. Remember that these quality measures are unreliable approximations and should not be taken overly seriously.

There are only two values really - good (keep) and bad (reject), with a region in between that is a tossup.

[Sukhi Singh]

Thanks Istvan, makes sense. Either q>25 or q<10 Cheers

Answer 2

[Sukhi Singh]

As pointed out correctly by Brad, it is due to the change from five-parameter quality model to six-parameter quality model.

From the tech note

"Why did we move to the 6-predictor model? Although the 5-predictor model was very good at predicting quality, the 6-predictor model is more accurate and enables us to accurately predict the high percentage of Q40 data that was missed with the 5-predictor model. The new model is also faster and provides Quality scores after around cycle 11 in read 2 of paired-end reads (compared to around cycle 25 with the previous model)."cycle 25 with the previous model)."

Read it here http://dl.dropbox.com/u/6634542/RTA_Quality_Predictors_TechNote.pdf

So this graph is correct and makes sense now.

Sukhi