[Brian Bushnell]

Thanks for the clarification, and thanks for sharing your data!

I did some mapping of the first 16m reads, and generated the following graphs:



The "Other" category refers to soft-clipped bases, which is very high in this case because PhiX is small so many of the reads went off the end (*Considering these reads have been adapter-trimmed, I have no idea what is being sequenced past the ends of the PhiX genome; it might be interesting to investigate). Overall the average error rate is below 1% but above 0.1% across the read. Read 2 has a higher-than-expected insertion rate in the first half of the read. Oddly, R2 has some Ns only in the first half, and R1 has some Ns only in the second half. Unlike other platforms, the error rate for R2 seems fairly flat across the read.


This is a different way of looking at the same data.


The quality accuracy graph indicates that again the Q-scores are binned, and like NextSeq V1, they are highly inflated. Over 70% of the bases were assigned Q41, but the average observed quality for Q41 bases was actually Q31.


The insert size distribution is fairly interesting for a couple reasons. It looks like the platform can probably handle inserts over 450bp fairly well; there were some short inserts, but they did not overwhelmingly out-compete the long ones. But the flat distribution of the short-insert tail is odd.

Lastly, it's worth noting that around 83% of the reads mapped to the reference with no mismatches or indels.

For comparison, I've attached the mhist of a 2x150bp HS2500 run (not on PhiX), below. To me the HS2500 looks better, but not drastically better, in terms of error rates.

[pmiguel]

Hi Pmiguel,

the basic procedure looks like:
- 5 ul of library (2 nM to 3 nM including PhiX)
- add 5 ul 0,1 N NaOH
- add 5 ul Tris (200mM)
- add 35 ul Enzyme Master Mix
- load all 50 ul onto cBot

Ah, that's very interesting. They were finally forced to kick that ridiculous 50X dilution/neutralization step to the curb.

So you cluster at 200-300 pM. About 10-15x what we use on our HiSeq2500.

--
Phillip

[DNATECH]

Thanks a lot for the detailed analysis Brian.
Lutz

[pmiguel]

The insert size distribution is fairly interesting for a couple reasons. It looks like the platform can probably handle inserts over 450bp fairly well; there were some short inserts, but they did not overwhelmingly out-compete the long ones. But the flat distribution of the short-insert tail is odd.

About the size distribution of the library vs. size distribution of the amplicons that actually cluster. I created a thread some years ago about a somewhat extreme sample clustered on the MiSeq:

http://seqanswers.com/forums/showthread.php?t=20839

The 4th post in the thread, I actually converted the mass-based/log-linear plot results from the Agilent bioanalyzer chip to a linear, molecule-based plot. This way it can be directly compared to the insert sizes found by mapping the reads-pairs back to the genome from which they came.

The result showed that the shorter amplicons must have clustered preferentially. Really preferentially.

To me this has always suggested there must be some sort of competition for clustering that favors shorter amplicons.

At the much higher clustering concentrations using for the 3000/4000 this process may be exacerbated.

--
Phillip

[Brian Bushnell]

The 4th post in the thread, I actually converted the mass-based/log-linear plot results from the Agilent bioanalyzer chip to a linear, molecule-based plot. This way it can be directly compared to the insert sizes found by mapping the reads-pairs back to the genome from which they came.

The result showed that the shorter amplicons must have clustered preferentially. Really preferentially.

To me this has always suggested there must be some sort of competition for clustering that favors shorter amplicons.

Impressive; I was under the impression that inserts much over 800bp simply would not bridge-amplify. Maybe we should try that approach! Anyway, rather than shorter molecules vastly out-competing longer molecules at all lengths, that could be a more of a case where the rates are fairly similar up to a point (1kbp?) after which longer molecules start failing to form clusters at all (even if there were no short molecules present). I'm just guessing, though.

[SNPsaurus]

Brian, when we were developing local assembly of paired-end RAD, we were surprised to see contigs of 1200 bp being assembled (see http://journals.plos.org/plosone/art...l.pone.0018561 figure 4), meaning that there must have been fragments of 1200 bp undergoing bridge amplification. We had to use a "triangle cut" in the gel size selection to over-represent the larger fragments, but they did bridge.

I think the size preference in the patterned flow cells could be because a small fragment could enter a well after a larger fragment but then outcompete the larger fragment to fill the well. Or in the diffusion kinetics?

[pmiguel]

Impressive; I was under the impression that inserts much over 800bp simply would not bridge-amplify. Maybe we should try that approach! Anyway, rather than shorter molecules vastly out-competing longer molecules at all lengths, that could be a more of a case where the rates are fairly similar up to a point (1kbp?) after which longer molecules start failing to form clusters at all (even if there were no short molecules present). I'm just guessing, though.

We did cluster at 1/2 the normal density, so that may have allowed the longer amplicons to form clusters where normally they would not have. Again, my natural inclination is to regard this as some sort of competition. Looking at plots of insert sizes and comparing them to the sizes of the input library it has always looked to me as if all the amplicons queued up by length and then all the shortest ones clustered. Okay, an exaggeration, but more-or-less fitting what one sees.

--
Phillip

[DNATECH]

The latest HiSeq3000 run (we did receive a few flowcells) did average 378 million clusters passing filter, per lane. All libraries were size selected.

One obvious part of the exclusion amplification as implemented is the very viscous enzyme mix. Probably the diffusion of the library fragments towards the flowcell is very much slowed down (requiring also higher library concentrations?) giving the molecule that arrives first the chance to become amplified and fill entire nanowells before a second one arrives (http://www.google.com/patents/WO2013188582A1?cl=en). The viscosity enhanced "drag" also could explain the stronger bias towards smaller inset size reads?
The high viscosity buffer together with high library concentrations and "RPA" amplification for the clustering process ("Recombinase Polymerase Amplification" ( http://www.twistdx.co.uk/our_technology/ )) might be sufficient for the Kinetic Exclusion Amplification on the nanowell flowcells? It seems to me that the other methods described in the patent might not be compatible with the old cBots (these can be used for the Hiseq3000/4000 clustering after a software upgrade)?