We recently acquired a NextSeq machine and are not very impressed with the data. I've uploaded a spreadsheet containing some of the statistics here:
The first tab is a HiSeq2000 2x150bp run. The insert size was below target, so I adapter-trimmed adapters before analyzing the data (no other preprocessing was run); and the HS2000 is not really spec'd to 2x150, so as you might imagine, the quality suffers toward the end. Regardless, it's pretty good. Looking at the mapping stats, 99.55% of the reads mapped, and overall 79.85% of the reads were error-free.
The next two tabs contain a couple of lanes of NextSeq bacterial sequence. Lane 1 generally seems to be the best, with quality dropping to a minimum at lane 4. But even for lane 1, only 96.47% of the reads mapped and 49.3% were perfect matches; by lane 4, 95.91% mapped and 38.91% were perfect. So the rate of reads with errors roughly tripled from HS2000 (which does not support 2x150bp runs) to NextSeq (which supposedly does), and as you can see on the "Average Quality by Position" and "Error Rate vs Read Position" graphs, the comparison would be brutal - an order of magnitude or more - if you consider 2x100bp reads. Also, if you look at the "Quality Score Accuracy" graph, the HS2000 quality scores are fairly accurate and typically underestimate quality, while the NextSeq ones are inaccurate and overestimate quality by about 10 dB (and are quantized), so you can't easily quality-trim the NextSeq data to improve it.
The "Library Uniqueness" graph, generated by sampling a kmer from each read and hashing it to see if it was seen before, is also very odd for NextSeq. It is wavy. The graph should monotonically decrease and any increase indicates a sudden error burst. So it seems maybe the period (~625000 reads) corresponds with an image frame, the clusters around the edges of the frame are blurry, as one might expect from low-quality or miscalibrated optics.
The Base Frequency vs Position graph is also interesting - NextSeq has a clear A/T ratio bias that is not present in HS data. The 3bp-wavelength sawtooth pattern probably has something to do with codon structure.
Does anyone else have data they'd like to share on NextSeq machines?
P.S. Command lines I used:
The first tab is a HiSeq2000 2x150bp run. The insert size was below target, so I adapter-trimmed adapters before analyzing the data (no other preprocessing was run); and the HS2000 is not really spec'd to 2x150, so as you might imagine, the quality suffers toward the end. Regardless, it's pretty good. Looking at the mapping stats, 99.55% of the reads mapped, and overall 79.85% of the reads were error-free.
The next two tabs contain a couple of lanes of NextSeq bacterial sequence. Lane 1 generally seems to be the best, with quality dropping to a minimum at lane 4. But even for lane 1, only 96.47% of the reads mapped and 49.3% were perfect matches; by lane 4, 95.91% mapped and 38.91% were perfect. So the rate of reads with errors roughly tripled from HS2000 (which does not support 2x150bp runs) to NextSeq (which supposedly does), and as you can see on the "Average Quality by Position" and "Error Rate vs Read Position" graphs, the comparison would be brutal - an order of magnitude or more - if you consider 2x100bp reads. Also, if you look at the "Quality Score Accuracy" graph, the HS2000 quality scores are fairly accurate and typically underestimate quality, while the NextSeq ones are inaccurate and overestimate quality by about 10 dB (and are quantized), so you can't easily quality-trim the NextSeq data to improve it.
The "Library Uniqueness" graph, generated by sampling a kmer from each read and hashing it to see if it was seen before, is also very odd for NextSeq. It is wavy. The graph should monotonically decrease and any increase indicates a sudden error burst. So it seems maybe the period (~625000 reads) corresponds with an image frame, the clusters around the edges of the frame are blurry, as one might expect from low-quality or miscalibrated optics.
The Base Frequency vs Position graph is also interesting - NextSeq has a clear A/T ratio bias that is not present in HS data. The 3bp-wavelength sawtooth pattern probably has something to do with codon structure.
Does anyone else have data they'd like to share on NextSeq machines?
P.S. Command lines I used:
Code:
bbcountunique.sh in=reads.fq.gz reads=100000000 out=uniqueness.txt bbduk.sh in=reads.fq.gz reads=4000000 ktrim=r k=25 hdist=1 mink=12 tbo tpe ref=nextera.fa,truseq.fa out=ktrimmed.fq.gz ow bbmap.sh in=ktrimmed.fq.gz reads=4000000 mhist=mhist.txt ihist=ihist.txt bhist=bhist.txt idhist=idhist.txt ehist=ehist.txt qhist=qhist.txt idbins=200 qahist=qahist.txt aqhist=aqhist.txt indelhist=indelhist.txt gchist=gchist.txt bbmerge.sh in=ktrimmed.fq.gz reads=4000000 ihist=ihist_merge.txt
Comment