Wednesday, February 07, 2018

Oxford Nanopore Outlook 2018

I'm behind on these posts.  My usual foibles were largely responsible for a while, but then I had the major (and sad) family issue that has kept me off balance for  two weeks. Someday I may write about that, but for now back to the major sequencing vendors.  Though with Oxford Nanopore, the problem is where to start?  But now is the time to get moving, both since Oxford's Clive Brown will be webcasting an update on Thursday and I'll be at AGBT next week and expect to be busy with news flow from that event. Clive's webcast is titled "sub-$1000 human genomes on Nanopore (and other goodies for H1 2018), so expect quite a casserole of tempting updates.  Certainly it is enough to get me to try to be fully mentally awake at 6 am, something that does not come naturally.
I realize the adage about glass houses and stone throwing applies here, but Oxford IMHO is pursuing too many different aspects of their technology and risking not devoting enough energy to the most key ones.   But what are those key issues?  Your opinion will probably depend on how you use the Oxford technology.  But I'll pick mine below but also cover (and mark) the ones I feel are distractions.


Oxford is flying high from the full publication of the human genome sequencing paper (which of course has been available as a preprint for many months), as well as the publication a bit earlier of their direct RNA sequencing paper (which also had a preprint).  The paper received very wide media attention and the fact that it boasts the first haplotyped assembly of the Major Histocompatibility Complex was often noted.

I wrote a critical piece on the idea of PromethION over a year ago, and mostly stand behind it.  That said, it does appear that Oxford is getting closer to expanding the PEAP early access program and even starting to make noises about commercial launch.  At least one outside group has tweeted out multiple views of PromethION in action.  But it is also clear that some groups haven't gotten a steady supply of flowcells to meet their sequencing desires.

One thing that isn't clear is how many flowcells advance PromethION users are able to run in parallel.  As with MinION, the yields that Clive Brown is reporting internal at Oxford -- around 120Gbases -- are not approached in the field, where the current record appears to be 40Gb.  For MinION, a few users have finally crept into the 20Gbase range and approached Clive's 25+Gbase runs, but many users are mired down at 5Gbases or lower.  Clive recently tweeted out what appeared to be a shelf loaded with PromethION boxes, but whether flowcell supply has reliably ramped up remains a question.

A working PromethION system is certainly a serious challenge to Pacific Biosciences.  PacBio certainly has a large lead on placing boxes -- they just announced 10 going to BGI -- but if your application can tolerate Oxford's data quality (as I covered in my PacBio piece) then the productivity of PromethION will be astounding.

PromethION won't be for the casual user: the minimum flowcell order will be 24 and that sets you back over $50K.  To get the cost of a human genome down to under $1K, assuming yields far beyond that seen in the field, will require committing to over $1M in flowcells -- the best pricing requires a $1.8M commitment and has "subject to availability" next to it (after all, that is for nearly 3000 flowcells).

In addition to the yield variation and the persistent data quality issues -- the human paper used Illumina polishing -- new users of PromethION may be disappointed if they cannot obtain top-of-line library size distributions.  That's still clearly an art, and the longest reads seem to not come from libraries with the highest read N50s.


At the New York meeting in December ONT indicated that the scaled-down SmidgION cell phone sequencer and the adaptor to run on MinIONs, or Flongle, are likely to show up late in 2018.  That's too bad, because to me Flongle is likely to be revolutionary in gaining acceptance of the platform.  SmidgION will certainly be a media hit, but it is Flongle that will really bring in users and usage.

Why do I apportion things that way?  SmidgION has serious cool factor -- sequencing on a smartphone.  It could be very useful for field applications -- but it doesn't solve the in-field sample and library prep problems.  Those are supposed to be solved by Zumbador, which is another prototype I'll cover below.  But that's the problem, without Zumbador SmidgION is a rather incremental improvement, but one requiring major development efforts.

On the other hand, the flowcells for SmidgION, with what is presumably a relatively simple adaptor called Flongle, are potentially revolutionary.  If ONT can succeed in launching the "crumpet design", with the expensive electronics separate from the cheap flowcell, then Flongles might be really, really inexpensive.  That would be a boon for fun niche markets such as education, but more importantly it could really enable MinION to be used in clinical and industrial settings.  The current capacity of the MinION overshoots many applications, and while flowcell washing and/or barcoding can mitigate this, really inexpensive flowcells would be a far better solution in many cases.  Library prep might become the dominant cost, but purely detection assay formats that don't involve much prep -- think a Nanopore formatting of Nanostring assays -- could mean a few dollars per sample to get it ready for use.

If industrial groups really did start going gung-ho for Flongle, then the box that would be really needed would be a GridFlongle -- something that runs many downsized flowcells in parallel.  Presumably such a device -- as my suggested moniker implies -- could be evolved from the GridION.  Ideally this would have a geometry similar to 96-well plates, but if that's too hard there are plenty of workarounds such as variable-spanning automated pipettors.  And I certainly think there is a much bigger market to run huge numbers of very small assays in parallel than to run large numbers of huge sequencing runs, which brings me to...


GridION is definitely off on a good pace, helped by a mostly evolutionary hardware design that leveraged the working MinION system.  The number of service providers using the platform remains small, though just after I remarked there was no North American provider one launched, DeNovo Genomics in Kansas.  

An interesting idea for GridION, hinted at by Clive Brown and also thrown out by others, would be a modification to accept PromethION flowcells.  As a strategy from the start this would have been far more cautious and IMHO more successful.  Now, the exact setup of such could take many forms.  Users would probably prefer one that could be reconfigured regularly to accept either flowcell class (or any of three, once the SmidgION/Flongle flowcells launch), but Oxford has hinted that future GridION designs might not accept MinION flowcells.  One could also imagine a fixed, mixed configuration -- you could order your machine to have five stations with your choice of a mix flowcell types but no field ability to change a station's flowcell class.


I'll confess that if I were running things, the Zumbador field preparation unit would probably be put on a very far back burner.  It certainly has serious cool factor and will significant impact scientifically by further enabling field operations, but it is also a very difficult challenge.  Or perhaps the best path would have been to license out the basic design to some other group that would develop it.  Oxford has shown very minimal appetite for partnering on anything around their platform, going it alone on kits.  


VolTRAX seems to be another poor performer, stuck in a very long beta phase after an even longer alpha phase.  Apparently the only kit available is still just the rapid 1D, which is so simple on its own that most users don't see added value from VolTRAX.  ONT has often tried to tout Rapid 1D on VolTRAX as less variable than manual libraries, but that claim doesn't seem to really hold attention.

Data Quality

Data quality remains an issue across the ONT platform.  Ryan Wick's spectacular preprint masquerading as a README file gives a nice overview of how this has gotten better with different basecallers.  For many applications in which long-range information is more important than base perfection, nanopore is doing well.  But the fact remains that for many applications polishing with Illumina is imperative, which can't sit well with ONT management given their visceral disdain for  Big I.  Oxford recently released a new read correction tool called Medaka that uses run-length encoding for homopolymers and is claimed to obtain similar results to Nanopolish with less compute, though at the moment Ryan Wick finds performing worse and crashing on some datasets.

It's worth noting also that 1D^2 chemistry seems to not have caught much fire.  It can't help that this isn't available in a rapid format.  Perhaps the major market for 1D^2 will be amplicons, since that is an application where taking a consensus of multiple reads can be problematic.

Nanopore has never taken up my suggestion of generating defined test articles to better assess performance.  BioBricks has recently put out a request for useful DNA to be synthesized and made freely available to the bio community, so I'm actually designing some to contribute.

Direct RNA

The direct RNA sequencing mode has certainly caught a lot of attention, I would expect will continue to grow in demand.  Note that direct RNA flowcells appear to not be in production for PromethION; this is a MinION/GridION specific application.

Kits and Other Miscellanea

Oxford will almost certainly keep pumping out new kits and applications. The "Cas Me if You Can" Cas9-driven capture is being tested by a few select users, as is a field version of the Rapid 1D kit.  Clive tweeted out that a single-cell sequencing kit is under development.  Given all the single cell solutions in the market, I think I'll tag this as another example of ONT insisting on forging their own path for the sake of forging their own path.

MinKNOW for the GridION has a much better design to my eyes (which are poor indeed, combining nearsightness, peculiar color response and a pressing need for bifocals).  In particular, the cacophony of different pore states, which confuses any novice, has been reduced to something close to a logical lifecycle of a pore (ready, sequencing, recovering, dead).  Rolling that redesign out for MinION users should be a priority. I have a bunch of other complaints about MinKNOW which I'll save for later, but perhaps Oxford will short-circuit those as well.

If there's one area that Oxford has skimped on, it is really solving the input sample question. What contaminants are trouble, which are minor nuisances and which don't matter at all?  That would, of course, be a spectacular use case for GridFlongle, as you can't very well multiplex samples that may be individually destroying flowcell performance.

There's also goodies promised but not yet delivered.  At London Calling, it seemed that flow cells that wouldn't require priming were in the final stages of being readied; these do not appear to have been launched.  Eliminating this would take away a nuisance step that can trip up novices.

Thursday morning should be interesting -- Clive is never dull -- and with a little diligence on my part I'll have something Thursday or Friday to compare these predictions against any new announcements.


David Eccles said...

1D² is unlikely to work well for amplicons because the complement detection won't be able to distinguish the same DNA sequence from a different one. See our chimeric reads paper (which you've demonstrably read at least once before) for an example of such a fast pore re-loading from an obviously different template DNA strand.

I still think ONT should concentrate primarily on the hardware and chemistry side of their device, and leave the software to the computer scientists. Make a great MinKNOW application and API for hooking into the raw signal data and making decisions about whether to sequence or boot strands, and leave the base calling for people who have more/different DNA to play around with, or lots of experience with signal processing. Now that Chiron v0.3 is out, and giving similar (if not better) results when compared with Albacore, I don't expect it'll take too long before the global software community has optimised base calling to be quicker than ONT's own partially-open-source callers.

One of ONT's many aces up their sleeves is the sampling rate of the hardware. It's capable of pushing out over five times the rate it does at the moment, which will give huge accuracy boosts once models are retrained.

James said...

"I realize the adage about glass houses and stone throwing applies here, but Oxford IMHO is pursuing too many different aspects of their technology and risking not devoting enough energy to the most key ones."

You hit the nail on the head here. ONT is still a very small company and I think they are spreading themselves too thin. Like most small companies they need to focus on a few small applications and grow from there. For example:

Improve minion reliability: From my experience there can be an order of magnitude difference between yields on minions for no noticeable reason. Having a 2GB run and 10GB run using the same protocol is unacceptable. If ONT can consistently ship minions that can get between 10-15GB of data that would be huge. This type of reliability is needed for production sequencing centers (which is the majority of the market).

Create a reliable analysis pipeline: We need fast and reliable analysis tools that work for a broad range of applications. ONT needs to focus on building production ready toolsets for their users to use

Improve accuracy: Either hardware or software improvements to solve the consensus accuracy limitations need to be the #1 priority. ONT only data peaks at best at QV30. There's no point of releasing a Promethion that can produce a $1000 genome if the accuracy is 99.8%. People will still use Illumina to sequence.

I really hope this doesn't become a case of great technology run to the ground by poor management. The minion and gidion have been out for a while now and at most ONT has shipped 10,000 minion. That's about $7million worth of revenue over the past few years. Illumina gets over $2.5billion/year and Pacbio is about $100million/year. Money talks and it's clearly not heading in ONT's direction. They're so focuses on hyping up random products than focusing on markets that they can actually compete in. The current state of the minion and gridion don't really make them appealing for any large scale projects. This is proven by the fact that only a handful of academics are buying them and no one is using them for large scale projects.