Poor read quality in GAII

[rakumar]

What factors can lead to poor read quality?
Can it be because of the library prep even if the enrichment was fine?
Is there any way to determine if the poor read quality was because of poor cluster generation?
Thanks!

[james hadfield]

It doesn't look that bad to me.

PhiX and Lane 3 are a little different at the end but you still have lots of Q30 data in both.

I'm not used to seeing the data represtnted quite in this format.

[rakumar]

Thank you James. These were lower than what I usually get so I was curious. I had multiplexed my samples with barcodes being the first 5 nucleotides read followed by a T (RNA seq data). Usually we don't have a problem, but this time we got only 23 million reads per lane instead of the 35 million we get in the s_3_sequence.txt file.
Our sequencing support specialist mentioned that the drop in sequence quality score might be because I didn't balance the barcodes properly. I do see that was the case, with 50% of the nucleotides at cycle 1 being A (See attached pdf). However, I still don't understand why the cluster detection program will have a problem with this. Do you know why this will be a problem?
Thanks!

[visivas]

I think 23 M from 35 M is not a drastic decrease. That said biases upfront can cause the cluster detection programs to behave awry. The cluster detection program depends on all four nucleotides to efficiently locate the cluster at the first 3-4 cycles. For example if you have 50% of A, then the other GCT images will look dark for the most part. The program will interpret this as a lower intensity cluster or some problem with the cluster formation itself.

Apart from this I would also check your initial intensities for lane 3 and how that drops over cycles. Seems to be a drastic drop in the last 10 cycles which is not common in more recent times.

[fkrueger]

Hi Rakumar,

I agree with James that the qualities look quite fine overall. What seems to happen is that quite likely a subset of reads will have B basecall qualities starting at different positions in the read, and this usually extends all the way to the end of the read. Like this it looks like the later cycles cause a quality drop, but in fact this is just caused by the subset of reads with BBBBBs (Phred score of 2) throughout (The whiskers in the quality plot go down to Phred 2). Just trim reads containing Bs and I am quite confident that the quality will then look like the PhiX lane.

There is a good chance that these quality values are a consequence of your indexing used, especially if some indexes are overrepresented compared to others. A possible explanation why you "lost" more clusters this time could be that you didn't really have 4 barcodes as expected, and the raw cluster density could have been higher this time than the last times. The closer clusters with the same barcode come together and the more data will be lost because clusters won't pass the PF step. We have done some work on this in the past, feel free to have a look here: http://seqanswers.com/forums/showthr...ack+processing.

[rakumar]

Thanks everyone for the comments. I just wanted to make sure that the decrease in the quality I am seeing is not because I did some steps of the library preps on the spri beads.
fkrueger, thanks for sharing your findings about the low diversity and data loss. One way to tackle this issue is to increase the number of samples multiplexed (and therefore increasing the barcodes diversity) right? So instead of doing two 12 plexed hiseq lanes, running of 24 plexed hi seq twice would be better?

[fkrueger]

Thanks everyone for the comments. I just wanted to make sure that the decrease in the quality I am seeing is not because I did some steps of the library preps on the spri beads.
fkrueger, thanks for sharing your findings about the low diversity and data loss. One way to tackle this issue is to increase the number of samples multiplexed (and therefore increasing the barcodes diversity) right? So instead of doing two 12 plexed hiseq lanes, running of 24 plexed hi seq twice would be better?

Not quite, according to our findings and a simulation of the cluster calliong process the problems are most sever if you have one, two or threee barcodes only. Anything above that is reasonably diverse to allow efficient cluster detection.