Can ONT Maintain Grip on Burgeoning Flowcell Herd?

A notion dawned on me when I was mentally planning my write-up of the Nanopore Community Meeting, but I decided to put off fleshing it out until a later date.  After a bit of procrastination plus a crush of other ideas, here it is: Oxford Nanopore flowcell lineup has been a bit complicated for a while, but it's probably going to get worse.  There's always been serious issues with the current level of complexity and it's hard to believe this will do anything but escalate.

One challenge with ONT is keeping up with their lingo; the other is filling in the spots they haven't.  I need to differentiate several aspects of flowcells and I'm not sure there's official nomenclatures, so bear with some definitions.  First, there is the form factor.  Currently ONT sells three: Flongle, MinION and PromethION (I'll just call the one MinION; of course it works in GridIONs as well). Clive's NCM talk touched on at least three more: there's the 96-well Plongle, the SmidgION that just plugs into a standard port and a redesigned MinION device with on-board sample prep capabilities.  Plus there's sequencing-on-a-VolTRAX, which isn't a typical flowcell but my arguments below apply to it.

Next you have the pore chemistry that goes on this.  Technically, this includes the motor proteins and other components, but we'll just go by the pore numbers.  ONT currently sells three different pore chemistries: R9.4.1, R10.3 and R9.5.  R9.5 is specialized for 1D^2 chemistry, which has never really seemed to gain much user share, and you actually have to special order it.  Mentioned by Clive were two more chemistries (well, technically he mentioned R10.3 replacing R10.0, but I'll skip that) -- there's the R9.6 to support the 2D-C method (which is almost certainly the death knell for 1D^2 and R9.5) and there's a future R11. Plus possible devices with both R10 and R11 on them  There were also some Clive asides suggesting that further iterations of R9 and R10 are in development as well. So a bunch of chemistries.

Then there's the underlying hardware around the pores -- the ASICs and new super-dense design Clive talked about at NCM.  So far ASICs have matched 1:1 with flowcell form factors, but it seems likely that a future MinION format device might have one of the new ASICs to reduce cost and eliminate the wash-and-return requirement (often observed in the breach, but it does exist).  I once hoped to have a simple grid of all this, but I've quite given up.  It hasn't quite gotten to four score and seven pores, but it feels like that's where they're heading.

Then there's library kits -- there's the ligation kits, the rapid kits and PCR libraries.  Plus the direct/native RNA kit.  Some applications even seem to care about which flowcell they're run on; getting some of the Cas9 capture schemes to work on Flongle apparently took extra doing.

Then there's promised boosters coming soon -- the new tethering scheme to enhance loading sensitivity, the ATP regeneration method to generate more data.

Then there's the different computational components: basecalling, polishing with Medaka and methylation analysis with Tombo.

So that's a lot of different aspects that most users would like to not worry about.  Well, for some you might want to choose -- I'm quite taken with the R10.0 data I've seen and its superior consensus quality.  But, that comes at a cost of lower overall throughput and lower sensitivity, plus the raw accuracy in my hands is actually not clearly better than R9.4 -- R10.0 has a much tighter distribution but R9.4 gives significantly more very high accuracy reads -- just good luck knowing which those are!

R9.4 is ONT's workhorse and they appear unlikely to change that anytime soon.  With the exception of 1D^2 and 2D-C, everything is supported on R9.4.  Any of the three current form factors, all of the library kits and every released or research-grade software tool supports R9.4.  A reasonable guess is that Plongle will start with R9.4.  The 2D-C is starting out with an R9 series pore.  And so forth.

In contrast, R10.3 has only minimal support so far.  You can use ligation libraries, but not rapid.  You can call bases and polish with Medaka, but not work on modifications.  R10.3 on MinION is shipping now, but PromethION is soon and Flongle is promised but not scheduled yet.  So there's a lot of applications where you might want the vastly superior consensus accuracy of R10.3, but that means giving up other things you want, such as Flongle, rapid libraries or methylation calling.

In the same vein, you can get everything in the original MinION flowcell format.  But PromethION is very much more limited and Flongle even more so.  Partly that's due to Flongle having trouble meeting the R9.4 demand. But the more specialized chemistries -- particularly 1D^2 and now 2D-C -- seem to only show up in the original form factor.  For many applications having accurate reads in Flongle will be very desirable.

Product line complexity carries a manufacturing and distribution challenge for ONT, and I've frequently criticized their logistical prowess as I've often been frustrated with it in real life.  Since flowcells are a perishable product, its all the worse -- Oxford can't just build up a huge inventory and let it play out.  Clive briefly described some of the manufacturing innovations that ONT is working out -- apparently the key flowcell component will be effectively printed on a sheet as it goes between reels -- but having so many lines can't but complicate things.

For nanopore the complexity problem is exacerbated by the fact that the signals are so complicated and must be decoded with machine learning models which must be trained.  Change the ASIC or change the pore and everything must be retrained.  So ONT must generate sizable test sets for each permutation of ASIC and pore chemistry they decide to market, perform the compute intensive training and testing of models for that data, and push those models out to the user base.  Not to mention that in an ideal world there would be programs like Medaka and Tombo that could leverage data from different chemistries in a productive way.

On the library prep side, presumably some aspect of the engineering of R10 class pores explains why the rapid kits can't work with them.  Or perhaps the adapter is different enough it shifts the signal yet again.  Hopefully the new tethers and the ATP regeneration system will play nice with R10 pores, but we won't know for sure until those kits hit the market. 

Similarly, R10 seems ideal for 2D-C since the consensus accuracy is better (indicating less of the error is systematic than with R9), but it will take investment by Oxford to make this so. I really hope they do, but there's the risk that trying to support all of these will diffuse attention in a way that stimulates poor productivity.

Getting out of this dizzy whirlwind is now an option with the Q-Line, but of course it also means opting to be satisfied with today's status quo.  And that doesn't address the missing pore chemistry - form factor combinations which would be very useful in the markets Q-Line will be targeting, such as all the missing Flongle options, since Flongle is both easier to use correctly and has a megabases per unit cost metric that is more attractive (and sufficient megabases to be useful in many amplicon or similar assays).

Other than some minor pruning -- 1D^2 and R9.5 -- it's hard to see Oxford reducing this complexity much.  I suppose in some sense I'm arguing it two ways -- I worry about their ability to handle a large number of flowcell options but then I want many of them.  If I were in charge, I'd might try to switch the main company flagship to R10, but perhaps that is just more difficult from a wet-side perspective in interlocking with everything else. 

Perhaps the best solution is some carefully planned fact finding and survey of users to understand which components of the overall matrix will be in the most demand.  I would say that getting all library kits, or at least all DNA library kits, to work with all future flowcells should be a very high priority.  Medaka and Tombo working on all flowcells should be the equally high priority on the dry side (well, semi-dry -- there's the training and test data generation to be done). 

But the real challenge is figuring out the market wishes for the more specialized devices and chemistries.  For example, how do possible customers plan to use Plongle?  If it is in screening assays, then perhaps R9.4 is the only need.  But if users wish to use the 96 separate "flowwells"  to augment the (far too limited) multiplexing capacity to enable de novo sequencing of many separate samples, then there will be a demand for R10.3, R11, R10.3+R11 and even perhaps R9.6/2D-C on this platform.  Similarly, if SmidgION is mostly for hobbyists and educators, then perhaps only one chemistry is appropriate and R9.4 is appropriate since it is the most tolerant to low concentration libraries.  On the other hand, maybe SmidgION will be on point for dealing with the next incident like the current Wuhan outbreak, and so having high molecular accuracy will be important.  And of course not only does the Nanopore community have at least as many opinions on a question as there are options for that question, but there is also the need to anticipate the desires of potential customers that could be lured in with the correct product offerings.

So I don't envy either the complexity of dealing with this gemisch of a product line nor with trying to figure out what a varied and evolving marketplace might want it to be.  Plus they must wrestle all this while suffering the indignity of being kibitzed by the likes of me -- that shouldn't be wished on anyone.