This year, instead of touring London after the Nanopore confab I headed to Italy for the European Society of Human Genetics meeting. Upon hearing Lakmal Jayasinghe describe ONT's proteomics plans, I was debating what to do with my pre-LC piece on the peptide sequencing. Perhaps I could ask a restaurant owner in the shadow of the Duomo whether it would go better with risotto Milanese or perhaps as an addition to a minestrone. But with more thought, while I remain intrigued by ONT's concept, I still think many parts of what I wrote still stand up.
Whetting The Proteomics Appetite
Multiplex protein measurement recently has meant LC-MS (though technologies such as Luminex still exist). These are high capital cost instrument requiring highly skilled operator to run complex workflows. Oxford Nanopore's mission has long been to target such situations with their low cost, compact instruments and ideally generate experimental workflows that are simple.
If we look across "next gen proteomics" - a set of approaches that have resemblance or ties to next-generation sequencing - multiple new technologies have been coming online. These can be roughly grouped into two technology classes. First there are affinity reagent (antibody or aptamer)-based. On the market now or nearly there are Olink(owned by ThermoFisher), SomaLogic (partnered with Illumina), Alamar, Nomic, and Codetta. These use panels of DNA-tagged affinity reagents (in the case of aptamers, the tag is an inherent part of the affinity reagent) to bind to protein analytes of interest. Most work on some sort of sandwich principal - only if two different antibodies bind the same target. Off in a special category of its own is Pixelgen, which is designed to detect interacting proteins not individual proteins.
One key advantage of this approach is that the affinity reagents can be tuned and diluted to adjust for dynamic range - abundant protein targets have less binder than rare targets. With good affinity reagents, the specificity of the sandwich effect can be very high. Using polyclonal antibodies can reduce the risk of polymorphisms or post-translational modifications throwing off the assay by interfering with binding.
On the downside, one must first develop the panel of binding reagents and then consistently produce them - will users see the same sensitivity and specificity with this year's lot of reagent as they did with the lot at the beginning of their five-year study? All those reagents must be produced and mixed precisely. And there is always the possibility that an affinity reagent is sensitive to the context in which it is used and switching to a different one - a very different tissue preparation or perhaps material extracted from formalin-fixed, paraffin embedded (FFPE) clinical samples - will lead to poor performance. This is a knock routinely thrown at SomaLogic/Illumina aptamer approach, though I don't have the experience to know if it is fair criticism.
As an aside, there is exciting progress on using machine learning to design binding proteins for known structures. Perhaps in the not-too-distant future one will be able to order high quality, custom affinity reagents the same way one orders custom peptides or oligos. But that day isn't here yet.
Because such assays rely on pre-built, pre-validated panels of affinity reagents, such panels are only built for popular assays - which pretty much means "whatever is currently hot in human and mouse work". If you work in plants, bacteria, fungi, invertebrates or even most vertebrates, these proteomics technologies are unavailable to you.
The other general line of approach are peptide counting methods, many of which start looking like next-generation sequencing but its sequencing-by-degradation not sequencing-by-synthesis. QuantumSI has the only launched system, Illumina-funded Encodia may have recently gone poof, but there is a gaggle of hopefuls in this space including Portal, Glyphic, Erysion, and DreamPore. Nava Whiteford just posted on one more nanopore company to surface, Swiss-based Unomr. And while drafting this Nava spotted a non-nanopore protein sequencing company, PumpkinSeed. And of course Oxford Nanopore.
Peptide approaches main challenge is the dynamic range one: abundant proteins will generate a lot of peptides. There's also a size bias - large proteins generate more peptides than small proteins. Warp Drive Bio's platform had a component for identifying the binding partners of novel rapamycin-like molecules, and it's no accident that the first successful de-orphaning (and sadly, the only published one) involved a human protein called CEP250 - which stands for Centrosomal Protein, *250 Kilodaltons*. Yeah, that generated many peptides so it wasn't hard to spot a bunch of them in mass spec. Depending on your digestion reagent, some proteins may generate very few usefully diagnostic peptides - far too many short and non-unique and perhaps some huge peptides that may or may not work (in mass spec they can fly badly).
A Blended Approach
ONT just announced a hybrid approach - working with a really new startup (incorporated this year) SienaQuant. Choose one digestion fragment to represent protein & have antibody against it. The basic concept is not new; a company called Epitome Biosystems early this century. Epitome was going to feed earlier mid-plex immunoassays such as Luminex, but failed to gain traction in the market and vanished, its binding reagent catalog slurped up by Millipore. With the SienaQuant/Epitome approach, these peptide-directed antibodies are used to capture their targets from a sample, which are then read with nanopore scheme - nanopore readout just gives a "squiggle" which you match to library of such squiggles.
On the plus side, the single targeted peptide levels the playing field between large and small proteins. Antibody doses in the capture cocktail can be used to bring the capture rates into a similar register - if you can reliably predict the expression range of a given protein and dose accordingly. With ONT as the readout, closely related but distinct peptides from related proteins - say, two different MAP kinases - may generate different squiggles. Reading out actual peptides is also an antidote to the increasingly recognized problem that far too many antibodies aren't nearly as specific as users had deluded themselves into thinking. And it may also be possible to distinguish sequence variants or post-translational modifications by their squiggle, giving much more information than Epitome-into-Luminex. And ONT is claiming they have sensitiviy in the hundreds of zeptomoles, which is a regime I'm not used to thinking of. 100 zeptomoles is 6 million molecules (are there 100 zeptomoles of zeppoles in all of Italy at any moment?).
But, you still need a developed and validated catalog of antibodies - so again non-human, non-mouse work is locked out with the possible exception of sufficiently cross-reactive peptides from other mammals. So maybe the mouse antibodies work on rat and perhaps some dog peptides are sufficiently conserved for cross-reactivity, but zebrafish or Drosophila will have slim pickings of such cross-reactivity. Some proteins of interest may not have great choices of peptides. And depending on the antibodies, sequence variants or post-translational modifications might sufficiently interfere with binding as to interfere with capture - or at least throw off the quantitation accuracy.
Garnishing Proteins with Peptide Tags
ONT has a second scheme for recombinant work & synthetic biology - make peptide barcodes. So here ONT would have a design for a peptide sequence they know they can differentiate the squiggle of that barcode for every other barcode they design. Include the coding sequence for this barcode into your protein with appropriate protease sites and affinity tags and you can easily enrich for barcodes and feed them into the nanopore sequencer. With this one might perform an experiment such as tag a whole bunch of single-chain antibodies expressed in your favorite microbial host, inject them into a mouse bearing a xenograft tumor, dissect out the tumor, extract barcodes and count them.
Versus existing tagging systems such as Hibit, ONT offers the opportunity for enormous numbers of tags - they are already building a 1000 tag set. The one catch is that these are protein barcodes going onto proteins you'd like to measure in vivo - there's always a significant risk of the barcode itself or where you place it causing biological effects such as in vivo cleavage or mislocalization or aggregration or any of the multitude of molecular effects in a complex biological milleau . That could all be context-dependent as well Which is another reason to have a bevy of barcodes - one might label the same protein in parallel with different barcodes and then average across all those barcodes to get a final measurement. Of course with time one can build up a set of high confidence barcodes for a given application - these are known to behave well in serum, these for human nuclear expression, these for human cytoplasmic and so forth.
Whole Proteins Are On A Future Menu
Longer term, ONT wants to run entire proteins. Few details were proffered on this during the ONT session, but Jeff Nivala's talk at the end of the conference gave an update on his approach. Analyte proteins are fused to the protein ClpX. These are driven through the pore by electrophoretic force, then ClpX pulls the protein back through the pore to generate a signal. This can be repeated - the protein "flossed" - to generate better signal. But as Jeff Nivala mentioned in his talk, flossing alone may not be a sure route to high accuracy amino acid sequencing. The challenge is that the pore is too long, with too many amino acids interacting with the pore at any given moment. Couple that to the the 20 standard amino acids - or worse, the entire spectrum of possible moieties present in a peptide chain after post-translational modification or insertion of non-standard amino acids - and it's a really tough challenge. By engineering shorter pore proteins, it might be possible to reduce this exponential complexity.
Will ONT Serve Success?
Will ONT be successful in this space? Nobody has a good crystal ball on this. The whole field is emerging and the valuable applications aren't clear - we could see a repeat of ONT's challenge in the nucleic acids world of demonstrating fascinating technical success in many new applications but few if any of those applications driving overall financial success. ONT is a far more mature company and will likely avoid many execution mistakes they made with the early DNA program. But it's also a space where new competitors continue to come-and-go, as noted above.
On the antibody panel side, it will be interesting to see the lineup. ONT partner SienaQuant is nearly a tabula rasa, apparently incorporated this year and with a web domain purchased only this spring. But they are part of a more established company called SISCAPA that started in 2011 and specializes in using peptide-directed antibodies and stable isotope labeled peptides for mass spectrometry assays. But SISCAPA doesn't provide an antibody catalog online, so how deep it is an unknown.
Anyone signing up for the early access program should anticipate that the available antibodies will be in a very focused area and that there will always be targets of interest not covered. These days all the cool kids are doing immuno-oncology; in the past it would have been MAP kinase signaling pathways. Or perhaps something more aimed at serum proteomics - 35 years ago as an undergraduate for a course we profiled all the lipoproteins involved in blood lipid transport. That's an accessible set of more than a dozen proteins with a huge body of literature to compare to.
What Did The Chef Not Mention?
After hearing the presentations, I was starting to contemplate what I might ask some restaurant in Milan in terms of a favor - perhaps my earlier words would be nice as part of a risotto milanese or served cacio e pepe. If it really is as simple as a different family of pores in the same consumable, then maybe ONT shouldn't spin out the proteomics business. But while I'm still reconsidering some of what I wrote, there's still so much unrevealed on ONT's plans.
In particular, ONT didn't present much in the way of performance information. They did show very nice confusion matrices which demonstrated they can reliably distinguish a number of peptides based only on squiggles. They also showed that the electrical signals show interesting deviations due to post-translational modifications. And as noted above, they are claiming detection sensitivity for the antibody scheme in the zeptomolar range.
So what didn't they show? For one thing, overall throughput. They've chosen a pore - and ONT must have an amazing library of different pores - which gives Q20 for reading the DNA tags (which should give great options for sample barcoding). But the speed they will operate at was not mentioned. Similarly, the lifespan of the pores -- does yanking peptides through them cause them to break - wasn't discussed.
Perhaps of even greater interest is what is the chemistry for attaching the DNA tags and how efficiently does it work? That in turn would drive what input amounts of peptides are required. Coupling the tags could also be tricky since there are amino and carboxyl side chains in proteins. Are these ever tagged with DNA? Are such branched structures a problem for the pores? If some peptides have only one tag attached, are those problems? Are any post-translational modifications problematic? Reducing disulfide bonds would seem essential, but what about isopeptide linkages (ubiquitin and its ilk) or extremely bulky things like extensive N-glycosylation? And what does the workflow look like?
What nanopore-unfriendly substances will transmit through from sample to flowcell? Extracting proteins and digesting to peptides is a well-established set of methods, but so is DNA extraction. Plant genome sequencing talks at London Calling often have discussed mysterious nanopore inhibitors that leaked through DNA extractions. Hopefully trying to sequence a McNugget won't kill the flowcell either; why chicken DNA is so hard on pores still seems to be a bit of a mystery.
And I still remain convinced that the peptide proteomics market will have very modest overlap with ONT's existing customer base. Even in the human genomics space most labs are not engaging in proteomics and if they are the odds that a limited antibody catalog will overlap with a given lab's research interests are poor - and it certainly won't help with solving difficult rare disease cases since those nearly by definition involve understudied proteins that won't be in the catalog. The protein tags are interesting and might appeal to the current hot field of machine learning antibody design, but again those aren't the people who currently come to London Calling. So ONT really needs to start a push into entirely new spaces of academics and startups and that will take an entirely new mindset and therefore likely an entirely new set of ONT employees inserting themselves into the space.
So not preparing a bloggers feast just yet. But as ONT's protein program unfolds, I won't rule one out in the future - ONT may have the last laugh. And I'm always up for tiramisu, which is a dish best served cold.
No comments:
Post a Comment