At the top, new CEO Van Parys was well received opening and closing each day, and even handled cleanly the solo really hot potato thrown into his lap (more on that below). On Day 2 founding CEO Gordon Sanghera was warmly introduced from the crowd and received a long and hearty round of applause from the attendees.
Research Hardware: P48 Slips Away Quietly
ONT made it clear that for the near term they will be selling four instruments. MinION Mk D and GridION fit the smaller flowcells; PromethION P2 integrated (aka P2i) and P24 for the large flowcells. So the PromethION P48 disappeared quietly; death by omission. While the longer range tech talk spoke vaguely of solutions for pushing the performance boundaries both lower (and less expensively) and higher, no hardware to do so was announced.
The Live Lounge area had one very obvious change. In past years there were many tables out, each focused around a product area. This year there was just one table with existing products (photo above), and this got no post Tech Talk augmentation with new toys - much like last year. No new hardware was announced so the refresh point is a bit moot. But the single table was very noticeable - but with only 4 hardware types it isn't hard to fit them on one table
The still passionate P2 Solo community made yet another pitch for saving their machine, with a tartly worded question in the Tech Talk Q&A which was promptly tossed to Van Parys, who acknowledged the fervor but reiterated the company line that support costs for P2 Solo were out-of-line with its perceived value. As sharp as the question was, it was nowhere near the white hot level of a message I was shown from a private chat, authored by someone who I once saw speak at a Nanopore event. That described intense rage at the decision and saying it made the writer look favorably on BGI's CycloneSeq competing nanopore platform.
A change in hardware sales that occurred somewhat quietly in the past that I still don't like. ONT used to post all their pricing on the Nanopore Store website. Sure, it was often confusing with the different bundling schemes, but the pricing was there. Now it's not for anything other than MinION. So when, as just happened, somebody asks me last minute during a holiday weekend (and immediately after I was in the midst of a huge fraction of the ONT salesforce!) if the price they've put in a grant proposal is correct for GridION, and I think "that must be either PromethION or a gigantic bundle of consumables included", I can't check anything. Hope that bogus price doesn't kill the grant!
Adaptive Sampling Comes in Focus, Bearing Digital Panels
Adaptive sampling, the mode in which information early in a read is used to determine whether to continue the read or spit it out to wait for a faster chance at something interesting, has always been interesting technically but not always clear how much the company saw it as a commercial winner. Now it is clear ONT is leaning hard into the technology. The approach has gone thru multiple nomenclatures - I liked Clive Brown's original "read-until" but apparently it confused many people - and now we have a bifurcation. The technology is "adaptive sampling" but the products are digital panels. Versus other solutions in the industry - custom PCR panels, hybridization capture - the pitch is that digital panels require no further experimental work.
Digital panels are apparently as simple to attempt now as supplying a BED file of target regions and a FASTA of the reference genome. I say attempt, as something not discussed either on the main stage or in a Live Lounge presentation is that the success of the approach is dependent on what fraction of the genome is a region of interest and the read length distribution of the library. Unfortunately, about the only person who really grasps this in detail seems to be Matt Loose, and nobody has yet trained an AI model on his brain. So for design-it-yourself there still isn't a good guide to expectations.
As presented several (3? AI search isn't ) London Callings ago, even the reads rejected as being outside any region of interest provide useful data - they can be distilled for copy number information and small variant calls. These can be used, as presented by Sissel Juul, for polygenic risk scores or methylation risk scores.
An interesting tweak that is finally available was mentioned in the Live Lounge but not the plenary - regions of interest can have directionality. I wanted this in my one adaptive sampling experiment because I was trying to sequence across a breakpoint, so reads in flanking regions were only interesting if they pointed towards the junction.
ONT has already released one methylation focused digital panel and two germline panels. The latter consist of one targeting pharmacogenomics and the other for hereditary cancer. Interestingly, only the methylation panel is rated for GridION; the other two require PromethION. Another PromethION-class tumor somatic methylation profiling digital panel is on its way. A hematological cancer panel for methylation profiling is also imminent. Interesting, this is intended to not be barcoded, to ensure correct sampling mapping. But while timing information was given, the flowcell type was not. I'm guessing MinION/GridION given the single sample, but this omission from the slides is a bit irritating.
ONT has discovered that the reversing voltage spike for adaptive sampling can cause issues for basecalling, lowering Q-scores. By monitoring membrane only channels - channels that never got a nanopore and were previously completely ignored as uninformative - a noise profile can be built and subtracted from the productive channels - much as noise cancelling headphones sample the noise outside to clean up what one hears when wearing them.
Methylation Patterns
Methylation-focused digital panels point to the rising value of obtaining methylation data directly from native long read technologies. Short reads have had various specialized library preparation methods (Illumina's 5-base perhaps to become the most important in short order), but they do require these extra lab steps and there are varying degrees of data degradation from them. In the worst case - or perhaps should say most common case - full methylation+variant calling requires making two libraries, one standard and one treated to reveal methylation. And most of these methods pick up only 5-methylcytosine and not the somewhat rarer and more enigmatic 5-hydroxymethyl cytosine. Further library prep trickery can reveal 5hmC on short reads. But why go to so much trouble, when ONT reads all of this with no special library preps?
One of the outside presentations noted an interesting case of leukemia in which trying to find any known fusion or driving point mutation failed, but the methylation signature was clearly in the neighborhood of a known leukemia type and this finding drove a successful treatment of the pediatric patient.
In the tech talk, an interesting case was presented in which a father, their son, and their daughter all had a deletion in the Prader-Willi syndrome region, but only the two children exhibited the syndrome. Methylation analysis showed that the deletion was on the father's inactive copy - this region of the genome is known to be imprinted in a parent-of-origin manner - but the children each had the deletion on their active copy.
Several other external speakers described more examples of methylation signatures being informative and of efforts to build large catalogs of methylation signatures and tools to recognize these signatures. There's still work to be done, but it seems like "watch this space" won't last long and soon methylation signature analysis will be routine for both rare disease germline work as well somatic diseases such as cancer.
Q-Line: GridION or Bust
ONT Q-Line for regulated environments is an aspect of the product line that tends to escape my attention; I'm not working in regulated markets and ONT doesn't tend to spend a much presentation time on it. But it is an important area for them and for users in this space, offering a 3 year refuge from platform updates.
In previous years ONT had discussed extending Q-Line to PromethION; no sign of such in this year's plan. Which means several of the digital panel assays won't be coming to Q-Line.
Basecalling: One Model to Rule Them All
A bit of information that leaked in advance was that the newest release of the Dorado basecaller would not have a super accuracy (SUP) mode. I learned this from someone in the clinical space who was quite agitated by this seemingly sudden loss of a model they had grown to depend on. I failed to track them down later to see if how this resolved assuaged their anxiety, but I'm guessing it might. As ONT appears to have made a major technical advance, designing a new class of underlying model which delivers accuracy near that of the prior SUP model while requiring compute in like with the previous high accuracy (HAC) model.
This is important in many ways. SUP couldn't be used for adaptive sampling, as it couldn't keep up with the basecalling - by the time you decided to keep or reject a strand it would be too late. Even P24s could get backlogged on calling in SUP mode.
The new SUP model - well, it's really the only model going forward - also improves basecalling in short tandem repeats, an area that prior models were weak in. All the modifications are delivered with accuracy similar to the prior model
Dorado also now includes a small variant caller which outperforms the previous standard, Clair3. As mentioned in the presentation, the separation of basecalling from variant calling is likely to lose information which could assist in variant calling; Dorado small variant caller is an attempt to bring these into a common tool that performs a more holistic process.
Direct RNA Multiplexing!
ONT is definitely not downplaying direct RNA sequencing - the talks and posters were perhaps light on direct RNA but that doesn't reflect company confidence in it. Finally, direct RNA is being multiplexed with 24 barcodes! Same motor, same model - but more samples per flowcell!
It will be interesting to see how the community uses this. Direct RNA is still much more limited in throughput than cDNA approaches because the motor protein driving molecules through the pore operates at such a much lower speed than the DNA motors. So it probably won't make sense to try to barcode whole human transcriptomes, unless it would be to run the same pool across multiple flowcells to average out flowcell variability.
But for method development, this is surely a major advance. Or for RNA therapeutic QC, if great read depth isn't required. Or perhaps direct RNA sequencing of bacterial rRNA for looking for modified bases. Maybe applications that were simply blocked by the prior need to either devote a flowcell per sample or go through washing steps.
cDNA Kit Tweaks
On the cDNA front, a new kit has been released with a new reverse transcriptase recommendation which pushes the read length distribution farther out - double the median.
PromethION Plus Flowcells
New PromethION+ flowcells rolling out which aim to eliminate the nuclease wash kit usage. Graham Hall walked through an example where one customer with 10K flowcells used washing, whereas another of similar usage levels used the High Throughput Barcoding kit to pool multiple samples but then run them over multiple flowcells. This averages out the variability in yield per flowcell - in the ideal case the last few budgeted flowcells can be omitted to save cost if their predecessors truly delivered - but both approaches are problematic. Batching for the multi flowcell approach is likely to lengthen turnaround times; nuclease washing is a manual step that is undesirable in industrial settings. Wash-and-reload also carries risks of loading the wrong sample on a flowcell - there aren't enough barcodes available for proper barcode rolling strategies and that is assuming barcoding in the first place.
With the new flowcells, which feature a new buffer, ONT is claiming higher yields per flowcell than was achieved with wash-and-reload. The new flowcells will be released in July to high throughput users, with an indeterminate rollout to the broader community.
ONT declared a new focus (perhaps a grossly overdue one) - including building machine learning models - to try to understand manufacturing variability and correct it. Ideally should not be required to pool multiple samples and then run pool over multiple flowcells to average out flowcell-to-flowcell yield variability.
EPI2ME Updates
ONT continues to expand their EPI2ME software and associated applications. The processing flowchart for human whole genomes already starts looking like a serious subway map - and more tools are imminent. Joining the school of ONT-developed tools will be three more fish: mahimahi for haplotype-resolved clinical tandem repeats, sillago for copy number, variant calling, and phasing within highly similar paralogs, and chinook for genotyping, phasing, and complex structural variants involving the pharmacogenomic relevant gene CYP2D6.
On the microbial side, ONT has a lab protocol plus EPI2ME processing scheme called NO-MISS. This uses bead beating as a universal cell cracker, yielding reads with a read length peak around 8kb. ONT claims this can routinely close microbial genomes. I'm a bit skeptical of some of the challenging polyketide synthase repeats we saw at Warp Drive Bio. I won't count against them the truly pathological cases such as the rRNA array in Saccharomyces, which to my knowledge has never been assembled - and is suspected to expand and contract dynamically.
ONT is also working on EPI2ME workflows specifically for biopharmaceutical production and cell therapy. An mRNA QC pipeline using direct RNA currently covers four features - sequence, intactness of transcripts, detection of the modified base N1-pseudouracil, and poly A tail length. They hope to layer in 5' cap detection and double stranded RNA detection as well. An adventitious virus detection scheme for human cell lines is under development.
Next: The Future
I think I've hit all the highlights - feel free to point out in comments anything I grossly skipped over. My next post will cover the back half of the Tech Talk, which covered long term advances in the platform.
No comments:
Post a Comment