More information: For personal (as in human genome) sequencing, I'm not entirely sure how useful this is... We already have a scaffold for the human genome (that would be Mr. Venter himself) and MAYBE you could get some haploid information out of it, but you'd worry that the base pair you care about for any given individual is in one of the wrong stretches.

For organism sequencing, I think this could be even worse than pacbio; there are modestly sized gaps in the sequence compared to relative to the known pseudomonas sequence; imputing the correct sequence when you have no available template is going to be a nightmare. It's easy to say "15% error" when you know what the 100% correct result is. But if you have no reference correct sequence and all of your parts are up to 15% wrong, the difficulty is compounded.

Ya the MinIon is more of a DNA sensor than a sequencer in its current incarnation, but awesome for portability. Super expensive per base though. If they can't improve the cost significantly along with the accuracy it will never be a competitive sequencer.

You're wrong about the PacBio though, the raw reads are random and cancel out easily with consensus. PacBio is the least biased and therefore the most consensus accurate next-gen sequencer available. See my other comments for some citations.

Naive idiot alert.. could you process the "same" DNA numerous times then take the mode in each case? Or are the errors essentially ones that would be repeated each time?

Actually PacBio does just that to get better a better error rate. They basically have X amount of sequencing that can be done. You can spend that X however you'd like. If you want a sequence that is X long, you'll have a higher error rate. If you want a chunk that is X/10 long, you can circularize it, and thus sequence it ten times. This gets you better accuracy with the redundancy. DNA is pretty robust.

In practice though, even with these "circular consensus sequencing" reads, the error model is significantly higher than other technologies.

One doesn't need CCS anymore, though it is still an option for shorter insert lengths. The errors are more random than other platforms, so they resolve easily with consensus. You just need 20x coverage or more, depending upon the application. With enough coverage you generally get better accuracy than any other platform because it's the least biased technology.

CCS still works really well if you want incredibly high accuracy. See this paper and figure 3 for Q90 quality reads: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3811116/

But most people don't need that for their applications. So just regular consensus using PacBio data alone is sufficient for excellent (Q60 or better) consensus accuracy: https://github.com/PacificBiosciences/GenomicConsensus/blob/...

I don't think they push CCS much anymore. The focus seems to be on generating long reads and then using Quiver to call the consensus (https://github.com/PacificBiosciences/GenomicConsensus/blob/...).

That's probably wise of them. The last time I looked at data from them, the CCS reads were still very noisy. It's better to rely on your strengths. In their case, it's long reads.

Yup. Just keeps getting better (https://pbs.twimg.com/media/BpUs_MAIYAEHb2o.png:large).

Is there an accuracy problem in PacBio's published de-novo assembly human genome: http://blog.pacificbiosciences.com/2014/02/data-release-54x-...

it was assembled de-novo (IE, without HuRef), and the accuracy stats aren't horrible.

PacBio is ultimately more accurate than Illumina given consensus. Illumina has some systemic bias so the ultimate accuracy tops out around Q40-Q50 (same errors just repeat themselves) and PacBio will get you closer to Q60-Q70 because the random errors resolve out. So it's the most accurate platform actually. This singuler MinIon read however was barely mappable to the reference, so it looks pretty poor quality. But without more reads we don't quite yet know what the systemic error profile will be though for it. That's why they're adding more pore types, they say, to get multiple systemic error profiles to try to cancel some out.

All depends on the application and the price per base. If it's cheap enough, you can average over more reads, even with higher error rates. The error rate should come down though. I guess the price will also come down a lot.

I think the single molecule approach has both advantages and disadvantages. Also I think that Illumina's technology in some sense also only reads parts of single molecules, and the averaging is done later in-silico, much as the nanopore would do also.