GridION X5 - The Sequel

A presentation by Clive Brown, CTO Oxford Nanopore Technologies; notes by David Eccles, 2017-Mar-15. See the presentation here.

Preface

• Some people watching this in Australia 3am [and David Eccles in New Zealand at 4am]
• This is Clive's once-yearly update, given shortly before the main conference in May. A few things will be saved for the meeting in London
• Company mission: enable anybody to sequence anything anywhere
• Publications / elsewhere demonstrate that people are already attempting to use the MinION technology in lots of places
• The technology is designed to be used anywhere, with very simple workflows.
• Main device (MinION) is portable, with minimal capital cost
• System is effectively real-time, or very close to real time
• System is demand driven -- can be used, put down and taken up again many times
• Read lengths are intrinsically long, limited only by sample prep
• Accuracy is pretty good, and improving
• Sequencer is good at cDNA; this aspect has probably been undersold
• This talk doesn't cover everything available. Clive Brown has given a few other previous previous talks
• The last google hangout at end of March [2016] is a good place to start

How nanopore sequencing works

• proten pore embedded in membrane
• array of membranes embedded in electrical sensor
• speed at which sensors measures things is orders of magnitude faster than CCDs
• after a second, have 1000b read, entirely available for analysis as the run proceeds
• can pipeline bioinformatics pipelines based on sequencing
• can make devices work together to a shared goal; methods continue
• a piece of DNA will give you a reproducable signal that can be decoded into bases
• currently running R9.4 on mem 10, motor E8 (phenomenally processive helicase)

decoding signals

• originally done by HMM
• explosion in past 2 years around neural networks
• now good experience decoding signals using NNs
• many methods in signal processing area that could be used
• current methods can learn local context over a window

MinION device

• At least ~4,000 MinIONs out now
• Aluminium device
• Inject liquid sample into system, pores devour sample in real time
• 512 channels running at once, can get at 450b/s/channel v. high throughput
• Mark Ib -- no significant improvements planned in MinION

Technology workflow

• Clive doesn't like slide: implies linear workflow
• Need DNA
• Variety of kits; simplest/snappiest uses transposome complex. On long DNA, can add in adapters; 5-10 mins
• Lengthiest preps (ligation) take about an hour; working to smooth out variation in sample prep
• Can take sample out, can flush sample out, can put more in, can muck with sample while it's in the system
• While running, run NN basecaller
• Calling can be done on the fly, or post-run
• Programmable feedback loop; feature to come back with a vengeance in future

PromethION; elephant in the room

• Cake-tin sized benchtop sequencer
• Different ASIC, a few thousand channels, in aggregate 144,000 channels
• Laid out to be pipette friendly
• Can put sample in flow cell offline and while running
• Bottom is compute module, can write data out at full-pelt to external storage
• Box was designed when running at 30b/s, probably can't handle a full run for real-time basecalling
• Running at full-pelt can produce a very large 233Gb per flow cell in 48h, 11Tb assuming 100% bandwidth utilisation, 3x a NovaSeq; right up there in that category of high-througphut sequencers
• No chance of falling behind; not going out of date before the box is delivered

On-demand sequencing on PromethION

• All kinds of workflow tricks to optimise pipelines
• No need to wait for sample, can run 3 samples, 1 sample, 48, or multiplex
• Can deal with lumpy demand
• Turn-around time basically limited by postage
• Can be shipping data back while running
• If not enough slots, can just buy more sequencers
• When fully deployed, will provide significant competitive advantage

PromethION Flow cell performance

• A few bad channels, but mostly green
• Numbers are high, yield numbers high, above the threshold for shipping

PromethION scaling

• Not novices for scaling; Gordon worked on Glucose blood strips
• Working to produce more flow cells

PromethION performance and yield

• Key problem to do with flow cell blocking
• Promethion typically 10Gb in 6 hours, aim >50Gb per flow cell
• Firmware updates for higher
• Can probably run for up to 4 days
• Software mature, run in house all the time, evolution of MinION and GridION software
• Control in a similar way to MinION, in paralle

Instrument shipping

• 1-2 per week, a bit slower than expected
• expect all backorder done by Q3 (original prediction was Q3 2016)
• Putting in software that lets ONT do remote firmware updates
• Ability to swap out hardware/software very quickly

PromethION Flow cells

• First shipping 3rd April to 12 sites
• a little bit of hand-holding, will ramp up rate of flow cell shipping after that
• Haven't had a single dropout from waiting list
• target headroom performance is so high that it will not go out of date

PromethION design change

• Compute module will not be able to keep up with 1,000 bases per second
• Need a bigger box, getting too tall
• Decided to move all of the computing into a separate box
• Compute module becomes a switch that lets us stream data to a compute room
• People who want to run multiple promethIONs can cable up and have processing elsewhere
• Most people put PromethIONs on UPSs anyway, might as well put a compute module there as well
• Can add in up to 80 TFLOPS of computing in compute module; can handle 1,000 b/s on a fully-running sequencer
• Will map consensus callers and assemblers in box
• Can use idle computing for local bioinformatics; e.g. containerised workflows
• Network requirements much simpler for a single compute box
• To be rolled out; upgrades included in the cost of purchase

What's inside the compute box

• Cramming in FPGAs on PCIe cards, easily 60Mb/s, 80 TFLOPs
• Allows to fully keep up with real-time basecalling on PromethION
• Can also implement own proper version of things like read-until

PromethION evolution

• Currently 24 flow cells enabled, can dump data to external disk
• Q4 add compute module to enable all 48 flow cells with local basecalling
• 2018 additional evolution, rolled into current purchasing

Base call acceleration

• NNs are fascinating things
• Engineers have been able to produce versions that run pretty efficiently on FPGAs
• Tend to get higher uplift than CPU/GPU; power cost is lower
• Possible to use stripped-down, mini FPGAs for smaller devices

Base call acceleration design

• Typically 1.8 events per base, very high requirements
• MinION at full output needs 200 GFLOPs
• writing OpenCL versions of base callers, unlikely to be as optimal as VHDL

Base call benchmarking

• Working closely with Intel
• MinION about 240kb/s
• Promethion about 65,000 kb/s
• Can only utilise around 10% of available CPU
• on GPUs, seem to only be able to use about 2% of GPUs; not a good performance payoff
• on i7-type CPUs, can process 200,000 bases/s, still only using ~10%
• on Intel Arria / Stratix, getting to 1M-7.5M called bases per second, using 60% of available processing power
• pretty confident that this is the way to go

PromethION base call implementation

• current generation 9Mb/s, almost a Tb per day currently
• second generation of cards up to 4Tb of real-time calling per day
• looked at accelerator for MinION... but you can do that
• coding in the background a dongle, either separate or intercalated between MinION and computer that will do local basecalling, and stream out reads to the computer

MinION compute requirements

• Cloud base calling currently - Clive's fault that ONT did that
• If you provide a safety net, people start to use it as a hammock
• Cloud base calling will be discontinued at 21st March
• MinKNOW now has integrated basecaller which will do it for you
• A good high-performance laptop will be fine
• Can just leave computer running, will keep basecalling after finish of run
• Provide binary base caller
• Writing to shared drive, another computer can do base calling
• Most people getting 3-10Gb
• Theoretical maximum 200kb/s, best internal about 100kb/s
• laptop CPU can do about 40kb/s base calling
• MinKNOW can deal with this

Accuracy / chemistry / algorithm

• On 1D, R9.4 base calling modal accuracy of just over 90%, maybe a bit higher
• old 2D system at 250b/s

Consensus accuracy

• Basic message: accuracy improving, data amenable to polishing
• Most errors now falling in / adjacent to homopolymers
• Will think about releasing optimised consensus callers in a reasonable timeframe

Homopolymers

• ONT's unfinished business is held up by detractors; fixed by ONT, then detractors move onto the next unfinished business
• Scrappie package, learnings are migrating into MinKNOW base calling

Novel base calling

• Working from raw data (more about that later)

Homopolymer

• Recently held up by competitor as a systematic flaw; just another obstacle to overcome (e.g. black knight in Monty Python)
• Scrappie doing fairly well, consensus calling of homopolymers can be done using scrappy output
• not as mature as other callers, but methods will only improve

Base calling from raw signal

• Clive hates event calling, has wanted to get rid of it
• Current base callers just take raw signal, output base calls
• Neural networks trained to optimally extract features from raw data, architecturally / conceptually better
• Accuracy improves, scales to faster sequencing speeds
• Can go straight from raw to FASTQ / SAM, recover about 80% of disk space
• left with compressible integers in FAST5
• Developer versions released Easter
• base calling should just get better and better

Base caller landscape

• Clive likes open development
• Albacore is the production basecaller; can be run offline; fully-supported
• Nanonet is a research base caller, available under open source, not supported
• Scrappie is the New Kid on Block, limited support, available to everyone shortly
• Standard workflow is MinKNOW + onward analysis
• Can intercalate other basecallers; preferred by power users

Other accessory tools

• Lamprey; file-watching wrapper for open-source base callers. Does what old cloud program did
• If writing to shared drive, or external compute, can use Lamprey for local-cloud-type thing
• A middle ground between basic and power users

Throughput

• Clive took his own blood (with help), sequenced it himself, gets 20G per flow cell
• Other customers typically between 3-11G; would be nice to get everyone up to 20-30Gb
• This still only represents about 20-30% of what is possible
• Lots of reasons, many to do with extraction; people not knowing how much DNA they're putting in
• Trying to reach into upstream workflow; focusing now on good sample prep

Software improvements to throughput

• DNA complex would occasionally wedge on top of pore; takes a few seconds, once in, won't come out
• If caught quick enough, can do a bit of read-until and flick complex back out of pore
• This is no longer the limiting factor

Read length

• Read length = fragment length
• If a pore is presented with megabase sequences, it will produce megabase reads
• Other systems will fail due to photodamage
• If you can figure out how to get molecules into the system, can produce reads
• Josh Quick / Nick Loman managed long reads (>750kb), largely by avoiding pipetting
• have accomplished N50s of 60/70kb
• Probably no limit; limit is what can be put in. Clive expects 7Mb sequence should be able to be done
• Some nutcases at ONT think you can do whole chromosomes
• Need to take what is being learned and make it easier for everybody

Upcoming improvements for throughput and sensitivity

• MinKNOW upgrade, improved unblocking
• Lifetime of flow cell improved by 50%, yield per flow cell and cost per base goes up concommitantly
• Working on releasing an official read-until
• Working group looking on samples people are looking at, looking at best library prep to give best output

Improving / replacing 2D

• Introduced in NY meeting; phasing out 2D sequencing, replaced with 1D^2
• 2D has always been a problem; strands covalently joined with hairpin
• Accuracy plots have quite a different accuracy for template/complement
• Bad structural effects that bugger-up the basecalled signal
• can't get speed above 250b/s

1D^2

• 1D prep
• As template strand is drawn in, other end gets closer and closer to pore
• Finish first molecule, other molecule is sitting nearby on membrane
• If second molecule hangs around long enough, it will be processed by the pore
• This occurs naturally about 1% of the time, no joining between template/complement
• Trick is to make the second molecule hang around longer

1D^2 consensus

• Accuracy much sharper
• 2 strands look like and behave like individual molecules

Traces

• Open pore current; drop as molecule goes into pore
• When first molecule is traversed, current goes back up to open pore
• Second molecule is complement of DNA
• With some trickery, can make second molecule hang around for longer
• 60% of data comes in template/complement pairs; expect that ONT can get that higher
• Can get very high 1D^2 yields at 450b/s

1D^2 accuracy

• Modal accuracy of 97/98%, a proportion are above 99%
• Long stretches of perfect data
• Algorithm is not fully optimal
• Would like modal accuracy to 99%
• Base-caller was not Scrappie, so at least has homopolymer issues
• Will need to change to R9.5 pore; better at capturing second signal
• Can still generate 1D reads
• consensus calling know what pairs are, helps polishing
• metagenomics may be more important to look at single molecules
• expect this will be forwardly-compatible with 1000 bases/s

1D^2 release

• To developers 27th March, developer kit + base caller
• general release to community 3rd May
• 2D kits discontinued on 5th May

New product; MinION well established

• Over 4,000; just started pushing into China, India, Japan
• workflow getting better, work to do on input material
• Aim to make more runnable
• MinION not licensed for service sequencing; makes sense to Clive & Spike, but not anybody else
• MinION is your personal sequencer

Offerings

• Huge performance gap between MinION and PromethION; bookending the space
• PromethION might be too large (only 10/20 HiSeq 10)

GridION

• Will make GridION X5 available
• the sequel, because it follows on from both MinION / GridION

What is GridION

• Original system that was proposed by Clive / ONT
• Designed around loading membranes in the lab; tore up design in 2011 to change to loading at ONT
• Concept is sound: large arrayable computers that can work together or individually on samples
• For a long time, GridION wasn't taken off website

GridION X5

• Bench-top format, a big MinION
• 5 individually-addressable flow cells
• Inside, taken PromethION developments and shrunk down
• FPGAs inside, real-time base calling for up to 1000b/s for 5 flow cells
• Everything is in the box
• Allows for small group-level or service sequencing

GridION Production

• All mature, in build
• Very highly-manufacturable design
• In the zone

GridION Pricing

• Two ways to buy: capital loaded, consumable loaded
• Capital commitment of $125k, flow cells $300
• Licensed for use as fee for service
• At 10G per flow cell, $30 per Gb; at 20G/flow cell $15/Gb
• Capital-free model $475 per flow cell with support fee
• $47 per Gb at 10Gb per flow cell, $24 per Gb at 20Gb per flow cell

Also Nanopore service certified

• Institute-wide service
• will run training and QC certification process; contact support
• enables you and customers to know that samples can be processed

Summary tables for 3 products

• Should be product info on website

Shop opens for GridION next week

• Expect to ship devices in May at the latest

Dates

• March
• 1D^2 with developers
• Lamprey developer release
• Cloud base calling discontinued on 21st
• April
• Transducer (MinKNOW HP fix) on 20th
• May
• Broader 1D^2 release (3rd May)
• GridION flow cells 15th
• 2D gone on the 5th

Lots of things Clive hasn't spoken about

• A lot covered in more detail in London
• Devlopment on targeted CAS9
• Massively improving array sensitivity
• CliveOME (replaced, with ultra-long reads)
• wants to see about enriching immunoglobulin regions
• wants centromere-spanning reads
• Zumbador looking really exciting
• has to become from a drop of blood to a genome
• Flongle / SmidgION
• Metrichor / Epi2Me; becoming completely separate from Nanopore

Questions

• Devices for service: yes, but not MiNION
• System suitable for 16s
• Error rate improving
• Guaranteed performance metrics: certification scheme will probably include this; should talk to Louisa
• Scrappie avialable for developer licenses, some in MinKNOW
• PromethION compute box can be made rack-mountable (part of roadmap, but need to see demand first)
• Switch box / old compute module can have different output options
• Optiomal temperature for sequencing - temp rises, sequencer runs quickly
  • probably no problem at 50 degrees, not sure about colder
• Lifetime for flow cells * up to 4 days
• How much DNA is needed to make 20Gb?
• answer will be forthcoming
• MinION is available now in China
• Preparation of 1Mb DNA: look at Nick Loman's blog; just be gentle
• MinION on 3rd April - can draw down latest pore or not
• same base-callers work on both pores; just rate of capture
• Consumable & Capital models are both available in parallel
• No need to implement PromethION technology on MinION
• expect to improve throughput to 1000 bases/s
• in principle 40Gb/day (60Gb/day theoretical max)
• complement distorts DNA signal, probably by stretching
• segmentation - no methods to deal with this yet, but nothing scary
• Plan for clinical MinION use? probably
• How close to routine culture-free WGS? Zumbador problem; probably pretty close
• can pop cells, tagment
• problem is sensitivity
• in principle, if they can get to the membrane, will see a lot of them
• trying to improve array sensitivity to deal with low inputs
• drop of blood / spit into tube, wait 10 mins
• Conversion of ONT terminology
• Clive used Molarity when he did molecular biology
• significant stumbling block for people
• PromethION service fee - chat with ONT service team
• PromethION based on R9.5 pores? No one has run PromethION/GridION in field
• they will only ever be 1D or 1D squared
• MinION accuracy for short fragments
• pore accuracy is independent of length
• aiming to remove capture advantage for short reads