A concern rippled through the Twittersphere. Among genome aficionados Roche has a bit of a reputation somewhat akin to Homer Simpson crossed with the Grim Reaper: a bumbler in genomics with the nasty habit of killing good technologies. In particular, Roche is blamed for the failure of 454 to rapidly march their technology at anywhere near the pace of Illumina and allowing Nimblegen to be marginalized in the array and capture markets. If you hold this view, then the fact that Roche didn't just buy out PacBio is of some, but limited, comfort. They won't be directly running the shop, but clearly Roche will have some pull.
A more general concern of mine is what this deal will do to PacBio's priorities for further developing the platform. Much of that will depend on precisely what sorts of diagnostics are developed, but in any case smart minds are now going to be thinking very specifically of clinical applications and spending less time on more research-oriented things. In some areas there may be happy overlap, but certainly PacBio will be focusing on informatics to serve their diagnostic concepts and meet regulatory requirements, which cannot be good for the research side. Still, going bust would be far worse, so the deal can't be seen as all bad.
What sorts of diagnostics? That was a big question. After re-reading the release, it finally dawned on me that the deal puts the joys and burdens of development on PacBio, with Roche serving in a sales and distribution role. Presumably Roche will have suggestions, but between the characteristics of the PacBio platform and their current interests, one might try to make some guesses.
The instrument and platform have some key characteristics. At around $800MK a machine, this isn't going to installed in every physician's office! Clearly a model of a small number of strategically located central laboratories is more likely. Strategically located might mean in major markets, or it could mean locating next to airport hubs. FedEx is a great way to pull samples in to a central site, but that might limit applications with extreme time sensitivity. I would expect any application to play to PacBio's strength: very long reads. Reading methylation directly might be interesting, but I can't off-hand think of a medical application that has been demonstrated.
One key question is whether applications will look more like amplicon sequencing or more like shotgun genomics. Amplicons would suggest further development of the Circular Consensus Sequencing (CCS) mode, whereas shotgun projects demand Continuous Long Reads (CLR). Some applications PacBio has discussed might fall in-between. For example, direct sequencing of viruses might look very CCS if the viral genomes are short, perhaps less than 4-5Kb, but more like CLR if longer than that. Applications might also influence further chemistry development: available chemistries have offered a trade-off between accuracy and read length.
Required read lengths will also need to be played off machine usage; an assay requiring 3 hour movies means at most 8 assays per day per machine. That would affect the amortization of the machine, which back-of-envelope suggests an amortization cost around $150 per 3 hour run ($800MK(365*2) -- no downtime & 100% usage -- neither realistic!).
In the amplicon space, what might be attractive opportunities? Ion Torrent offers a very fast workflow if fusion primers can be used, the error profile tolerated and required read lengths aren't long. MiSeq is going to be a strong choice for amplicons shorter than about 700 nucleotides. QIAGEN may soon launch their machine, which is shrouded in mystery but will probably perform amplicon sequencing more cheaply than Illumina (among other things). A wild card is GnuBio, which is proposing really cheap amplicons with approximately 1Kb read lengths.
That still leaves room for PacBio -- perhaps. One application that Roche has experience with is amplicon sequencing of MHC regions, for which long reads are really valuable. Hospital for Sick Children presented a PacBio pipeline for cancer amplicon sequencing, but that seems to be a poor fit; weak argument for long reads (particularly if working with the notoriously fragmented FFPE samples) and lots of players there. Ditto for most variation analysis. Sequencing through nasty simple repeats to pull out disease alleles is a natural fit,, but are there enough such medically actionable alleles to support such a test commercially? Amplicon sequencing to profile infectious agents might be a win, but will payers pay for it? As mentioned above, direct sequencing of viruses has some overlap with amplicon sequencing.
Nick Loman suggested by Twitter that perhaps fusion genes would be a possibility. That's certainly an interesting idea, as a number of targeted therapies in cancer hit fusion genes and there is a lot of heterogeneity here. But can PacBio really beat out cheaper short read technologies here? That remains to be proven. Fusion genes also can be a small fraction of the transcriptome, and in clinical cancer samples the cellular heterogeneity may dilute them even at the genomic level. Optimized capture techniques might work here, but that remains a technology in development for this particular platform.
Methylation is a wild card. I'm unaware of a good medical application, and in a mammalian system some way of sub-setting the genome would be required to meet the capacity of an RS II.
Could RS II be useful for characterizing long mRNA, with the size selection serving to match the sample complexity to the abilities of the sequencer? Perhaps, but what is the medical utility?
Whole genome sequencing of microbes is challenging to see being a big market anytime soon. There are some great demonstrations of this for biosurveillance and better understanding hospital-acquired infections and so forth, but is that enough to be market? And again, does PacBio's superior assembly capability translate into a premium? Or, could margins be fat enough to bury the cost difference between PacBio and short read WGS? Or the erratic flow of samples would mean that short read technologies couldn't get per-run costs down?
I won't be shocked if something entirely different shows up. I don't claim to have encyclopedic knowledge of this space, nor do I have any inside view of PacBio's thinking.
Going back to Roche, a key question is what sort of partner they will be. Good partners collaborate; a whole range of ill behaviors can make a partnership toxic. If Roche fails to advance tools coming out of PacBio towards the market, then the really big payoff will never happen. Roche could also cause problems by trying to micromanage or regularly sending conflicting signals as to what is desirable. I'm not saying they will be a bad partner, but as with human relationships company relationships are very challenging to execute well. Roche also appears to have some degree of exclusivity in the human diagnostics space; that could drive developers away who are on the fence as to platform choice for their assay.
(corrected 9/26: $800K per machine, not $800M! Thanks to Nick Loman for spotting this)
(corrected 9/28 that bogus price had a second instance I missed before!)
(corrected 9/26: $800K per machine, not $800M! Thanks to Nick Loman for spotting this)
(corrected 9/28 that bogus price had a second instance I missed before!)
Haha yeah I was going to comment on the extreme inflation of the RS machine price!
ReplyDeleteI have to believe that Roche has something in mind regarding applying the deal towards oncology clinical trial use and/or development of companion diagnostics for targeted cancer therapies.
Since PACB had long been stating that the future of their business model depended on the clinical market, once cannot lament the losses around the edges for research applications too much. As you said, this deal is a major lifeline providing an alternative to bankruptcy or a GNOM-like takeout.
Based on the GNOM buyout filings, I think Roche was one of the companies kicking the tires and looking to do some sort of deal with them first. And of course they had interest in ILMN too. Maybe they don't care so much which tenchology to partner with!
Thank you for another fascinating article. I was wondering about your statement about PacBio's systems improved ability at "Sequencing through nasty simple repeats to pull out disease alleles". My understanding is that the majority of issues with sequencing short repeats, such as dinucleotide microsatellites, is strand slippage as a result of polymerase infidelity. Could you comment on why the PacBio system would be better at combatting this problem?
ReplyDeletePacBio / their collaborators had a poster at AGBT13 showing the ability to read a long VNTR in a mucin gene, a VNTR that turns out to be polymorphic and one of the alleles leads to a disease state. It may have been related to this PLoS One paper
ReplyDeletePacBio has no PCR, so the opportunities for mischief such as slippage are reduced. Perhaps more importantly, any error should be independent, whereas a slippage event early in PCR leads to a jackpot.
If the VNTR is longer than the read length of a given technology, then you just can't measure it (though certainly careful calibration of PacBio would be in order using some known standards); this will make Moleculo less valuable for this application.
The paper using amplicon sequencing on the PacBio system has published: Genome Reference and Sequence Variation in the Large Repetitive Central Exon of Human MUC5AC
ReplyDeletePacBo has been developing a smaller version of the RS for some time. I wonder if that one will be used in this case...
ReplyDeleteIt sounds to me like repeat diseases such as fragile X are likely to be in their sights.
ReplyDeleteHi Keith,
ReplyDeleteFraX is a good example. CGG-Repeat. Plus, there you have the methylation, as there are (rare) males with full repeat expansion (more than 200 Repeats, often 800 and more) and still no methylation of the expanded allel (high functioning expanded males). And while most PCR protocolls these days are very good at amplifying these repeats, there is always this remaining little bit of nagging doubt especially in females homozygous for a normal allel, which is why some old-school clinical geneticists would still consider the good old Ed Southern J Mol Bio 1975 as gold standard in that case.....
And as to methylation in clinical diagnostics: You missed more than a decade of imprinting diseases! Angelman Syndrome, Prader-Willi Syndrome, Beckwith-Wiedemann Syndrome, Silver-Russel Syndrome etc. However, with methylation specific MLPA, these are easily and cheaply diagnosed, so not quite sure if that helps much for a 800k machine.....
More commercially interesting might be promoter methylation in somatic cancer analysis. If you have a system that gives you mutations and promoter hypermethylation in one run, that might be a very big advantage over any other current system.
Alway enjoy reading your blog here in Germany.
Lars