Advances in Genome Biology and Technology (AGBT)

*This transcript was generated using an AI-based transcription tool and may contain errors. It reflects the spoken content of a live webinar and has been lightly edited for clarity.

Great. Thank you, Gustav. And he pronounced my name correctly. That's awesome. So, it's off to a good start. Great, so I've been thinking about things a lot lately, reflecting on the past and one of the things that came to mind was the very first time I did sequencing. So, it was quite a long time ago. Sanger sequencing, manual, one base per lane, autorad, your P32 and all that. And I'm bringing this up because two things. One was I screwed up, made a mistake in the experiment and didn't get the gel transfer step right. And two, because I didn't want my boss to know about that - and Seattle team don't listen to this, this is not something to listen to - I came in the middle of the night to do the actual sequencing of it. This is the first time I'd ever done it before. And I'm sitting there doing the sequencing and thinking, man this is so cool. That somebody figured out how to do this, and I'd say the wonderment and the awe that I had at that moment, I think it's something I never forgot, never want to forget about, and something that I think that was borne very early on in my mind, something that I wanted to be able to do is to develop a new sequencing technology and have that impact. 

And so I’m honored to be here presenting at this conference. It's been a long time coming, 18 years since starting Stratos. So, I will thank everyone for coming, and also to say that I very much respect all the technologies that I've seen here, and all the things that people are doing. It's amazing, and I very much love hearing about all the ideas and things that people come up with. 

So SBX technology. So, like any sequencing technology, you have to be thinking about accuracy, throughput, read length, cost efficiency. Those are givens. I don't care if it's 1999 or 2029, you have to be thinking about those things. But for us, flexible operation was one of the things that, when Bob and I started meeting at Starbucks for six-hour sessions, was right at the top of the list. We envisioned the technology that can sequence up and down the throughput spectrum, on the same system. Four minutes, 40 minutes, four hour runs. We wanted to be able to do that. In fact, I've talked to people, I'm thinking 42nd runs for some of the runs that I've heard from people today. 

So that was top of mind, and we'll talk about it a bit through the presentation. This will be a faster version of the presentation we gave at the webinar, but it is recorded so you can listen to it get a little more detail. John Mannion covered a lot of the data structure stuff there as well, so I recommend people go look at that. 

So we'll touch on accuracy. The results I'm showing here for duplex whole genome sequencing, for HG001 reference sample. You get an idea of the duplex sequencing accuracy that we're getting, and we'll cover that a little bit. Throughput, we did some seven plex whole genome sequencing reference sample, one hour run. We got over 30x coverage in that one hour run, and that's about five billion duplex reads in one hour. So you're getting an idea of the kind of throughput we're able to get here. We'll talk about the trade-offs between duplex and simplex sequencing, in terms of read length and accuracy. There's trade-offs there, to give a little bit of an understanding about that.

Time to result. We're going to show really cool result, Sean will cover that for sample to VCF whole genome sequencing in less than seven hours, which is really, really cool. And a fun way to represent the flexibility of this technology. So something that we've wanted to do from day one. And then cost efficiency. We look at it as all elements of the technology you have to be thinking about. Whether it's the chemistry, the measurement, including reuse of the sensor module, those kind of things. The quality of the data you're seeing. Everything, every element of efficiency, you have to be considering when you're thinking about this. So top of mind, we are very comfortable with where we're going and going to be able to go on the cost efficiency. And one thing, you know, be nice to Gustav! I know everyone's asking these questions about price and everything, and just to know that we've had him wired up the last week. I think Josh had the idea to wire him up in case if it gets in his mind to even talk about pricing, you know that he's going to get a little bit of a jolt. So, you don't want to see that. I saw it during the test runs, it's not a pretty thing. So anyway, be nice to Gustav. 

OK, let's get into it. So, OK, what is this? This is the coming together of Stratos genomics chemistry, and the genia array technology. I like to think of two technologies that needed each other to bring out the best performance. We showed the systems yesterday, introduced those. You can go see them in the suite, but basically a separate synthesis system and separate sequencing system. As you can imagine, those coming together at some point in time as a single unit is definitely something on our mind. 

OK, so what's our approach? So, our strategy for efficiently sequencing DNA was to not sequence DNA. It was that simple. Instead, convert DNA using an enzymatic chemistry process into an expanded surrogate molecule to address the signal to noise challenges of single molecule measurement. This was our approach it involved really on the edge molecular protein engineering techniques to do this. No one had ever done anything like this before. And so that was our approach. But when you get on the other side of it, you'll see the problems that you solve on the measurement side, what you can do when you have a chemistry, when you’re measuring a molecule like this. 

OK, the building block for doing this is something called an expandable nucleotide triphosphate or X-NTP. Our thinking at the beginning was attach a tether onto this nucleotide at two positions. And we tried many positions, but alpha phosphate and heterocycle made a lot of sense in terms of the biochemistry that we wanted to do. And then have those separated by a cleavable bond, and we use a bridging nitrogen amidate linkage there, to essentially, once you've incorporated this X-NTP in the copying step, replication step, that you can then expand the structure and now you have a code that represents the base identity. 

As time went on, we added translocation control because we knew we were really wanting to go in a nanopore direction and this allows for clean, reproducible, synchronized modulation of the movement of this Xpandomer through a nanopore. And we also learned that most polymerases don't like X-NTP, so we built in enhancers to actually improve the incorporation of the X-NTPs. And the last thing is nomenclature. SSRTs are symmetrically-synthesized reporter tethers. That’s what we refer to in the structure. 

The other big piece to this is a polymerase. For us it was pretty clear, we needed something that had an open architecture. You see where those arrows are, you need somewhere for these massive tethers to go. So we engineered a polymerase that had that open architecture. It's a lesion bypass Dpo4 polymerase, not conventional, not what you would normally think. But we needed to have something that had somewhere for those tethers to go. And I can spend all day talking about these last two slides. My team could, and they're here and ask them a lot of questions that they'll answer the ones that they can. But just overall, we had a iterative process of surveilling, refining, combining and iterating thousands and thousands of polymerases directed towards being able to better incorporate X-NTPs, and ultimately we ended up with something that has about 10% amino acid change. 99.3% raw read accuracy for simplex. Now obviously the duplex is much, much higher accuracy than that, but that's our kind of raw read total accuracy. That includes the polymerase copying step, and measurement, that's all rolled up accuracy. We turned it into a processive polymerase, which it wasn't initially and strand-displacing. So a lot of big engineering that we were able to do there. 

The workflow. I showed the synthesis instrument below, or in the suite we've shown that. It's basically just liquid handling, guiding reagents to this fixed solid phase chip that runs through a pretty intuitive process of, you have an oligo attached to a surface, you hybridize a template to it, add that cocktail of reagents to make the Xpandomer,  cleave to expand, and then using an orthogonal cleavage, photo cleave to release. That's the process. It's pretty intuitive, and it's what we currently do, and what we'll do for any sequencing right now.

So a little bit about that translocation control. This is a big deal because again, this comes down to measurement efficiency. Everything is about efficiency. And so I show there, with the green arrow, that little triangle holding the Xpandomer in place. This allows us to get good statistical sampling of the measurement, and you can see the level, the G level there. You then do a quick voltage pulse, advance the Xpandomer to the next step, and you see what that level looks like and so on, and so forth. So, you get an idea. This is a way of modulating throughput. So, if you pulse faster, you can increase throughput. And there's a lot of buttons that you can keep pushing on this technology to keep pushing throughput up. So that's one of them.

So that box there that I'm showing, was a drawing that Bob and I drew in 2012. This is what we hoped to be able to do. We couldn't do it then, but that was what we were drawing up. This is an image from 2019, so it took us a while, but this is what we were able to actually do. And you get an idea that we hit it on that signal resolution of measurement. That's a single molecule trace that we show there. You get an idea of what homopolymers look like, that we have this deterministic movement of an Xpandomer through a pore. So you can actually use that to call homopolymers, and we’ll show data later on how that looks. Quite good, actually. And then that's the resolution, the signal resolution we get between the bases when you look at the level histogram. So we hit what we wanted. 

OK, now you're on one pore. That was the data on a single pore on an Axopatch. Now you take it to eight million, OK. This is where the Genia technology comes in. You see the eight million chip on a PCB there, and it's essentially combining microwells, electrodes, detection circuits, and the A to D conversion. And then you get a picture on the right showing a bilayer, holding the pore in, and showing an Xpandomer translocating. 

So the sequencer instrument also liquid handling, to delivery agents, to the sensor module which houses the chip. And it basically runs through a very, this is a simplified version of the sequencing process, but you initialize a run, you set up the run by forming the bilayer and inserting the pore. We do this for every single run. And this is an iterative process, as I said. Then you go in and sequence for however long you want to sequence, whether it's four minutes or 40 minutes, clean the system and then you can reuse. OK, that's the process. Very clean. When we've looked at this many, many times, no crossover samples. And it's a very efficient process. And a lot of people have tried to do the Jedi mind trick on, on how many reuses. This is something that we'll talk about more over the coming months, but something that still is in process here. But we're very, we feel very comfortable saying significant reuse on that. 

Another picture to show is the pore count. So, at time zero, we typically get around seven to 7.5 million pores at time zero. But you get an idea over time, you know pores drop off over time, and that's kind of a typical drop off that we would see. But all the throughput that we're generating here, all went through this type of drop off profile. So another point to optimize with the technology, OK.

So, this is a one hour run single pore on the eight million array. The top picture you see open channel, you see the Xpandomer traces for many, many Xpandomers running through over the hour. And you zoom in and see what a 1000mer Xpandomer looks like. And the signal resolution you get there. That gives you an idea when you look at all the reads in that top window, how much throughput we're getting. And then you zoom in even further, and you see those translated XP base calls. 

OK, flexible throughput setup. We'll talk about this a lot, but just to say that if you want to look at one sample, four samples, because the synthesizer does four syntheses at a time, however deeply you want to multiplex or not, we want to be able to support these workflows flexibly. 

The paper. So we did a seminal SBX paper taking the technology up through 2020. So we really wanted a foundational paper to say, this is how you make these molecules and so that's available now. It has finally come up for people to look at, encourage you to read it. About 60-70 pages, you can scan the code here and have a look at it. Download took a while, I haven’t actually been able to do it yet so. But anyway, so it's a good read and kind of gives you an idea of the things we had to do to make the technology work OK. 

So, on the library prep, or at least the comparison between duplex and simplex, you get the idea. Duplex sequencing is basically measuring both parent strands on a single Xpandomer molecule. So, we can get intramolecular consensus reads on that. So that's the basic of that. And it will essentially show data on somatic oncology, as well as some of the fast applications, Edwin will cover the somatic, and Sean on the rapid. And so we'll go through a little bit of that. 

Workflow, pretty simple workflow. And it's essentially ligating to create that hairpin Y adapter structure. I say simple. It took a while to get a very clean workflow, but it is something that is very automatable and robust. And so we developed that protocol as well as a linear amplification approach. And that was key, it really helps us out a lot on the homopolymer accuracy, and so we were able to do a linear amplification approach and then just feed it in through the normal workflow to make Xpandomers in sequence. So other than that, you really don't know the difference.

OK, a little bit about the data. So this is that 7-plex genome in a bottle run. One hour run, we got over five billion reads into one hour. You get an idea of the breakdown of the number of reads per sample. You get an idea of the insert length coverage in that 34 to 38x coverage, and then the F1 score. So very respectable F1 scores for both the machine learning approach that John and Chen's team have developed internally, as well as comparing to the deep variant F1 scores, which they were very helpful and, and looked at the data and they actually got better numbers than us. So, it's nice to see that validation of our data in someone else's hands. Q scores. Q39, 99.7% genome coverage at greater than 10x. 

OK. The next workflow is SBX-Fast, which is basically removing the amplification step, and then tightening up the library prep, and just reducing some of the synthesis time to try and go for a really fast turnaround time on sequencing. And then just a little bit of data on SBX-Fast. We ran these a little bit differently. We ran them as trios. But basically five, nearly five billion reads, same type of quality of results. In fact, even better because we didn't actually even do the amplification step here. So great results, something that we really only just started doing. 

The vision, I think of the workflow, we think we can get in that five and a half hour range from sample prep all the way through VCF variant calling. We're actually not too far off of that and like I said, we just got started, but Sean will go through that.

A little bit about the read structure details. As you can imagine, we see this is a kinetic process. So you're going to see ranges of length, either full length or some partial. We use all of that information. We use all the duplex information, but we also use those single simplex tails. Those are really helpful for mapping purposes as well. So we use all the information that we have, very efficient, and you see very high yielding approach.

Accuracy over length was another thing that we looked at. John put this up here and described it in the webinar. Pretty much straight line. We don't see any impact, I think nothing to phase here. This is just single molecules running through a nanopore. On the homopolymer side, we see for all homopolymer lengths less than 15, we see F1 scores greater than 99%. That's both for our internal algorithms and development, as well as the stuff at Google Deep Variant, so very respectable homopolymer accuracy there.

A little bit about the accelerated data processing pipeline. Essentially this is where we're moving towards. As you can imagine, that's one of the strengths of the technology is to be able to literally read and analyze the data as it comes off. So looking at base calling demux, intermolecular consensus being done online essentially in real time, and then mapping and alignment in near real time. So that's essentially the direction we're going. And again, this is a big strength of the technology to be able to get that data quickly and flexibly and be able to use it when you want it.

A little bit about the simplex side. On the right, there is pretty much just single stranded sequencing. So we're not getting that intermolecular consensus, but there's a lot of good applications that if you want Q20 type of reads and a really, really high throughput like probe-based methods, or RNA isoforms, and Aziz is going to cover both of those in his session. This is a great approach to do that. And just to give a little idea of the yields and throughputs on the left, we show a 10x Flex run that we did. So a one hour run and we got nearly 14 billion reads in one hour. And those are near perfect reads because we pulled out from the total run, which was quite a bit more than 14 billion, just to give an idea of the throughput of links in that range. 

And then on the right side, you get an idea of a little bit bigger library size, around 800mer. And essentially if you look at all reads greater than 400, we end up with nearly over four billion reads in one hour. Over that, in that nearly 700 base read length, 99.35 accuracy and about 2.7 terabases per hour. So pretty substantial throughput. And again, like I've said many times, only getting started. Most of this data, we think there's a lot more in the tank. So I don't want to say disappointed. We're really happy to get this data, but there's a lot more here. So I see every experiment that we do. So I can say there's a lot, a lot more here to be done.

OK, So the wheel here in terms of the applications, this is what's so exciting, having talked to all of you here and some of you here today and this week is getting ideas about what people want to do with this approach. I love it. And I want to be able to see where we can go into the future and things that we can leverage this flexible technology. It may take some learning of workflows how to better leverage the approach, but that's, that's great. I mean, I think this is the whole idea here is kind of change the way we think about sequencing.

So with that, I can say I'm very excited again and honored to be here to go over the technology. It's been a long time, a long time to get here. Had the benefit of working with some amazing people over the years. And as I’ve said many times, Roche has been extremely supportive in this endeavor to bring this technology forward. So, I want to thank everyone and I think, Edwin, I think you're up. So, thanks everyone.

*This transcript was generated using an AI-based transcription tool and may contain errors. It reflects the spoken content of a live webinar and has been lightly edited for clarity.

Sean Hofherr: Hello. Pretty packed house! So, really excited to talk about our experience at Broad Clinical Labs with the SBX technology. So Broad clinical labs, we’re kind of nested within Broad Institute. We sit at this really interesting intersection between genomic explorers, resource builders, and clinical translators. So, basically folks across the whole entire spectrum. So, who’s looking for answers and new technologies, those folks that are trying to build up large databases so that consortia have access to really like, a plethora of data. And then, how do we take that information and actually apply it to clinical problems?

Here are a couple of examples of some of the folks that we've been working with over the years, and they run the gambit. And you know, a lot of these names are familiar and I'm going to mostly be focusing on the clinical translator side for the purpose of this talk.

At Broad Institute, we also happen to be kind of uniquely suited to test out a bunch of different platforms. So, we typically get the platforms before anybody else, so when they're in alpha release. Groups will come to us, ask us to try it out, give some real-world experience. We have the ability to provide input in real time before the product gets launched, which is really cool.

In addition, we also have like quite a large set of sequencers that are being run at scale. So this is like our current production lab. We have a large set of NovaSeqXPluses, we have a large set of Revios, we have some Ultima systems and then we also have some mid-plex and smaller sequencers as well.

So when Roche came to us, I guess this was like in the fall, and we had the first conversation about SBX technology, we were really excited. So, a couple of the things that jumped out and Mark covered this in the very first slide. The first aspect is there's a lot of data that can get generated from the sequencer. So that's like really, really attractive. In addition, supposedly high accuracy. At the point we hadn't really tested anything and the data that we were seeing was coming directly from Roche, so, we had to take it with a grain of salt. And then the speed of the system was also like really, really interesting. So being able to run and sequence genomes as quickly as we're able to do it on the sequencer. That actually unlocks quite a lot of potential. And one of the biggest places that we immediately thought that we could unlock potential is actually using this technology for rapid whole genome in the NICU.

So, today in the NICU, and Stephen Kingsmore gave a talk yesterday, he gave actually two separate talks. This figure on the right-hand side actually comes from one of Stephen’s papers and his group, and Rady Children's, has been pushing the envelope on rapid genomes for the last probably decade or so. 

What they found, and this has been replicated in many other studies, is that there's a huge diagnostic rate in the NICU. And so, when every single child in the NICU ends up with a genome, you actually find answers in like greater than 50% of the cases. And even in the cases that you're not finding an answer for, there's actually information that gets transmitted that helps in the management of that child. 

And so that's really important for a couple of reasons. So, the first reason is obviously you're improving the care of somebody that is like super sick and they're young, really young children. So that's really important. And the other benefit is when you start introducing genome into the NICU, you end up with, those kids actually spend less time in the NICU. And so that means now there's all of a sudden more beds like these valuable beds in the NICU for other patients that are severely ill to come into the NICU. And then lastly, the really interesting thing is that like the health economics of this makes a lot of sense. From the hospital setting, the healthcare industry setting, because having an early diagnosis actually leads to shorter stays. Shorter stays means less money that the hospital is taking out of pocket to cover the cost of that patient's care. And it's also less money that insurers are paying. So overall, this is a really attractive target. It's been a really attractive target. It's gained a lot of interest recently and it's starting to pick up and become routine care.

So, when Roche first came to us, they had this assay and it worked and we were kind of looking at it. And it was able to sequence a set of genomes in a certain amount of time, which was very nice. We wanted to know, how could we actually like dial down each one of the steps? So, as Mark said, there's a lot of levers. So you could decrease library prep time, you could decrease Xpandomer synthesis time. The sequencing itself is a big dial. So, you could actually have shorter sequencing runs and maybe load less samples on, and so you can get it as fast as possible. And then actually, how do we get variant calling to be as expedited as possible? Each one of these steps together there's a lot of opportunity to really shave off a lot of time and get this as fast as possible.

So, we got the instrument installed in January. So our colleagues actually had a little bit of a head start on us. And this is mainly because we were moving our facility from one building into another with all those sequencers. You could imagine how much time that was. So, in January, after the holidays, we came back, we had to install the system and actually within like one day, we were generating data. And so this right here, I'm just showing some of the initial data that we ran on the sequencer. So everything I'm going to be showing was generated in our laboratory, by our people. All the analysis was performed in our hands, and all the data came from us. 

And so, you know, the coverage is definitely higher than one needs for a genome, but this was kind of like the first shot. And so, we were just getting our hands, to get used to the system, and actually start thinking of those levers and how do we dial them up and down. But overall, the F1 scores for SNVs and INDELs were extremely impressive. And so right away we started thinking that, you know, maybe this promise that Roche came to us with, I don't know, six months earlier actually has some like some reality to it, which is really exciting.

When we start looking through all of the variants and we see how performance is across different variant types, interestingly, you know, this is a nanopore system and as you can see here, homopolymer performance is fantastic across precision and recall. And so this is like really, really impressive to come out of this system with such good homopolymer accuracy.

In addition, HighGC, LowGC, really good performance, really strong overall. Similar patterns for INDELs. I'm not going to go and kind of point out the differences, but in general the performance and profile is really, really solid for these variants.

And so, benchmarking with HG001 through HG004 is it's great, it's good, everybody's got to do it. But we wanted to know, are we able to find causative variants in individuals that have been diagnosed with genetic conditions? And so, we started testing some different variant types, so, cases with different variant types. The first set of samples I'm going to be showing are coming from just Coriell cell lines. Most of these are, are actual patients that were impacted with conditions that would be typically diagnosed in the neonatal setting. 

This first case is a case patient has propionic acidemia, it’s an inborn error of metabolism. These patients, actually in the first couple of days of life, if undiagnosed and untreated, even with treatment, they typically have like hyperammonemia episodes and they crash metabolically really early. And so, these kids do come to the NICU. And so, you can imagine just being able to have an answer for these patients could actually influence the care, because dietary management is a part of the treatment for these patients. And so, this is a patient, this patient has a single base deletion in the PCCA gene and we were actually able to detect it without any issue. This patient is homozygous. This is part of a trio from Coriell, so we do see it as heterozygous in both parents, and I'll show that data in a little bit. 

We wanted to look at performance across copy number variants. So we don't really just care about SNVs and INDELs. We also wanted to know how are we able to pick up copy number variants. This is a patient also Coriell cell line Duchenne Muscular Dystrophy. This particular patient has a 28kb duplication that actually like duplicates a couple of the exons in the DMD gene. It leads to Duchenne Muscular Dystrophy, genes also associated with Becker Muscular Dystrophy.

We were able to pick this variant up without any issues using the platform. Going away from the standard traditional types of variants, we're also interested in STR. So repeat expansions are a big interest for us. And so, the first patient that we tested, so that once again Coriell cell line, but we looked at a patient that had a full mutation in FMR1, so clinically affected with Fragile X. This patient we were able to run and detect the variant in this case.

These were not the only cases that we actually ran. We looked through a whole set of Coriell cell lines with different variant types. So you can see the first cluster, all small variants, they were all detected. The next set are all copy number variants also detected. And then the last set are repeat expansions with full mutations. And so, in each of these cases, the causative variant was identified. 

And so just thinking about the continuum, Coriell, great. But how does this actually work with patients? And so Broad Clinical Labs, we’re CAP accredited, CLIA laboratory also New York State accredited and licensed, so we have patient samples in house that have been diagnosed in our laboratory. And when we looked at those samples and we reran them, these are patients that have been consented to have additional testing performed with residual samples. You're able to see here that this set of samples, we were able to detect the causative variant that we actually reported in our laboratory for each one of these patients, which is really cool.

It's OK to be able to do this and see the variant in a VCF file or an IGV, but that's not really how we analyze cases in the real world in the clinical setting. So typically we take a phenotype driven approach and it goes into a tertiary analysis platform. This is an image from Fabric Genomics, Fabric enterprise platform. And what you see is that this is that PCCA, propionic acidemia trio and the causative variant was able to be actually prioritized using the AI. And this was without any training. So, we ran it as if it was a short read genome, threw the data in without actually even telling Fabric, and it worked perfectly without any issue and was able to find the causative variant. And this is what it would have looked like if we took this through a report. Just note this was not used for clinical purposes. We were just testing out to see how we could do it.

So, we got a little bit ambitious as we were going through this process and we thought, we are probably going to be really fast. And when you start looking, there's actually been a couple of Guinness records. So, Kingsmore had the Guinness record in 2018 for 19.5 hours, and this was using Illumina short read sequencing. Later he was able to cut it down, his group was, to 13.5 hours. Really, really impressive. In 2022, Euan Ashley's group using ONT technology was able to actually get a genome sequenced in five hours and two minutes, which is crazy fast. And actually going from patient to diagnosis, in a little bit over 7 hours. 

And so we actually like saw this and thought it was a great opportunity. And just without any training, or really much effort, we were able to get it down into six hours and 25 minutes. And so, we've only had this instrument for about six weeks. And so really the potential is huge and we're really excited to see how low we can go.

This was the work of a lot of people. Here are the main players that have been involved in this project for the last couple of months, and that's it. Thank you.

*This transcript was generated using an AI-based transcription tool and may contain errors. It reflects the spoken content of a live webinar and has been lightly edited for clarity.

Edwin Cuppen: Thank you Mark for warming up the room. Yeah, so the biggest question of course is for Gustav that everyone has in his mind. But the other question of course is also, how does all of this work in the wild? So, what does it work in a routine lab, and  what can you already deliver with that?

And I'm very excited to be presenting for the first time the work that we did in the past, well, less than six months. First a brief introduction on Hartwig Medical Foundation, that is a nonprofit organization and we’re actually doing routine whole genome sequencing based cancer diagnostics for various hospitals in the Netherlands. Besides that, we've actually developed open source analysis software to basically analyze this genomics data to drive data driven decision support tools in the hospital for oncology.

So these tools are used for two reasons. One is to generate a patient report with therapy guidance that goes back to the hospital. And the other part, with the consent of the patient, is to collect this information and bridge it with clinical metadata. And that is to basically build a Learning Care Cycle where with the data of the patients of today, we improve care for tomorrow's patients. So that's the objective of our foundation.

We do this whole process under ISO accreditation, and our average turnaround time is about six days. What is hampering us to get towards routine implementation, of course it’s new technology and it's a lot of data, but typically we're hampered by two main things. And one is basically this is too costly and it takes still too long. So, this is why we're always excited to explore new technologies to see if we can address these challenges. 

So we are very happy that we could team up with Roche on their new sequencing technology. This is a bit of timeline of the beta test that we did. You basically see the prototype equipment there, not as nice and shiny, but nice orange, which is a Dutch national color for the Xpandomer generation. So basically in October, November. So, this is installed in September 2024. So we started off with some test runs and optimization of our software, because I do want to stress that we try to be as independent as possible from Roche. Especially for the data interpretation, because we feel every new technology needs to be assessed under high scrutiny to see whether it's on par with what we're able to do already in our controlled clinical accredited environment.

Actually we did the first prediction run in December, and actually what I will be talking about is work and data that we actually generated at the end of January, which is done on 30 patients and some COLO duplicates. Ewart and Joris, I want to acknowledge them specifically because they really got the equipment very quickly working and under control in the lab. And also, Thomas and Peter, very instrumental as well to basically adapt our software to be able to work with data coming from the new sequencing equipment.

I don't have to say anything about this whole process. Mark already explained that. I will talk about 30 tumor normal pairs and two COLO control duplicate. Be aware, we do tumor normal pairs and we sequence actually per sample, per patient we sequence four genome equivalent. So it is a whole experiment and compares in total more than 120, 30x genome equivalents. So that was spread over eight, four hour runs that was done on a single chip and the sequencing was done over a couple of days to generate this data.

This was still done in a manual way, but Ewart and Joris also succeeded in automating this whole protocol and going from sample to data analysis or sorry, to data production, on four samples that could be done in 24 hours in an automated setup on Biomek. So meaning this is a versatile protocol that can be easily automated.

OK, what do we then get? Let's walk you through the different steps from raw data to the complete analysis. This is basically what you typically get, so what you see in these graphs is the tumor normal data and this aggregate information of the 32 samples. So on average, we've got 150 gigabases of data for normal, and 410 gigabases for tumor.

The spread of coverage is very good, meaning that there's not much variation between samples. This coverage, this is duplex data, but it's derived from two different setups. Mark just mentioned that full duplexes and partial duplexes in our setup. On average the duplex rate, the full duplex rate is about 70%, and 50 for these full duplexes is about 190 base pairs.

The partial duplexes have a slightly longer insert length, but you see that the termination of the duplex part is sort of random over the stem and that basically results in a duplex, average duplex length N50 and 110 base pairs. If you then convert it to mapped coverage of these reads. So that was the raw data, this is now the mapped data. We basically see that we get about 105x coverage for the tumor, and 47x coverage for the normal sample.

For the rest of the presentation, I will keep comparing this with our production data for these samples. So these are patients that were also analyzed on our inner production setup, that is currently running a Novaseq6000 and these patients had identical tumor coverage and slightly lower normal coverage.

OK, let's just look at coverage. GC biases. I think, don't look at the height of the curves, just coverage differences. But I think you should focus on the left and the right of these curves where you basically see that there is differences in GC biases where you basically see SBX coverage dropping off earlier than Illumina at low GC, but at high GC the performance of SBX is much better than Illumina.

So this results of course in dropouts in which you can nicely see in IGV views, you see dropouts in SBX in AT-rich regions, but the same thing happening for Lumina in GC-rich regions. If you look at the overall numbers the coverage, genome wide, is not much different. There is of course a second effect and that's the reads are longer. Compared to 2x 150 you have 300 base pairs to map something in a genome, and also you have a mate which is slightly further away from the other sequence, the other mate, so you get lower mappability due to the shorter reads. 

You see that in this in the white parts, the white reads have a map quality of zero, so they can also match somewhere else. You actually see that this region in this screenshot here already has more challenges in Illumina data as well, but you see that due to the shorter reads this region start dropping out completely with SBX data. However, if you look on where this is happening, and it's mainly outside the genome in a bottle, high confidence regions. So, if you correct for that and you just look at high confidence regions, you basically see almost identical performance for both platforms in terms of coverage.

OK. That's coverage, that's the basics, that's where you start. The next step is basically variant calling. We did this with our package of tools. So this is a package that we use in routine diagnostics. And it's not just variant calling it, there's also HLA typing, viral detection, all these types of things that can be relevant for the clinic, and for a cancer patient. You see the features here, you can use all the different types of data as input from FFPE, from panel etcetera. But one of the things how we designed this software is also to make it platform agnostic. So it should run with every platform. And the key tool in this is REDUX, and that's a pipeline component that deals with platform and sample specific characteristics and that feeds into the small and structural variant calling tools. And we basically took, well, took us a few months basically to optimize this tool to also work with the SBX data which has its own characteristics. And the rest of the tools of this pipeline were not modified, meaning they should all work once you have the REDUX tool properly working for this SBX data. 

So what this REDUX specifically does, it actually determines empirically the error rates to enable this high sensitive somatic SNV and INDEL detection. So that’s not the machine learning that you do once, it does that for every sample that comes in. So an FFPE sample has a different error profile and that's taken into account for the variant calling. And this is basically what it looks like for SNVs. It's trinucleotide context error rate that we determine based on a few percent of the genome that we survey for errors compared to normal reads. And this actually allows us also to calculate the empirical error rate from the different platforms.

On the SBX platform there is no different quality bins like there is for Illumina, but you have simplex and duplex reads. And if you apply this tool, so this is completely independent of any of the tools that Roche uses, we actually find the Phred score for simplex, for SNVs of 30, and for the duplex reads it goes up to 50. And if you basically compare the two platforms and consider all the data, we actually do see that the empirical platform SNV error rates are lower than on the Illumina platform.

Doesn't matter, this 1.7-fold or whatever way you calculate it. I'm not going to argue about that or make pleas about that. It's eventually. I think everything above Q30 is basically suited for the way we do variant calling in this for somatic variant identification.

OK, now basically what did we find? This is basically a COLO sample and basically here you see the challenge that you have with sematic variant calling. It's not a bin of zero, 50%, 100% variant allele frequency. No, it goes all over the range from zero to 100% and basically at the bottom of the graph on the left corner, that's basically where noise starts meeting real signals. Actually it's mixing, it's not even meeting each other. It's basically a mixed part of signal. So, it's difficult to interpret these graphs. Basically what you see in this case with this the raw data, you see high degree of shared variants.

For where do the private variants come from? Well about 20% is along these axes, no coverage in one or the other platform. That explains already a big chunk of the private. And the other part is basically if you look at the private variants and the variant allele frequency in the other platform, you basically see that they're all very low variant allele frequency. So we're actually looking at stochastic effects around just meeting the threshold for variant calling. So a simple thing to get this out of the equation is just to make a cut off of 10% variant allele frequency and then do the comparison. And then you basically see that you get sort of two identical numbers with the sort of symmetrical private profile, again around new cutoff that we defined. We believe these private variants are real variants when you look at their mutation spectrum.

This is about patients. You see the same pattern and you see very high concordance between a number of variants that are called in these patients with very low rates of private variants across these 30 patients aggregated. INDELS, same story, we basically see well, this is the COLO again. Please note that there is only very few INDELS in COLO. This is the melanoma sample, but the results are very clear that you don't get high rates of noise because of different specific error profiles.

Patients, same story for the INDELs. A lot of TF variants, and very few privates. The only privates that we do see are happening mainly at T homopolymers, which you basically see here. T homopolymers have much more abundant than C homopolymers in the genome. You basically see, above length of six or seven, you start seeing more Illumina private variants. We don't know if they're false negatives in SBX or false positives in Illumina. I just want to make a note that we don't use PCR Free Library prep in this setup, so these could also very well be PCR artifacts. 

Structure variance, copy number alterations. This is still work in progress. This is the first screenshot on COLO829. You basically see the inner circles that the copy number profiles are highly similar from both platforms. Structure variance is also looking quite promising, but we need a bit more tool optimization to get this completely done for all the patients as well.

What is the impact of private variants? Actually, if you nail this down, we see private variants across these 30 patients. But if you start looking at how many are in exoms and how many are actually actionable genes, clinically relevant genes, there are only 5 variants left, and they're spread over the two technologies. And the reason for missing them in one or the other technology is actually quite a variable, as you can see on this slide. And we basically come to the main conclusion that no clinical impact, there was actually no clinical impact due to platform differences.

I'll skip over the summary because this would be repeating what I just said and just come to the end conclusion that we feel that that's the SBX provides a very promising alternative platform for rapid whole genome sequencing based cancer diagnostics. Thank you for your attention.

*This transcript was generated using an AI-based transcription tool and may contain errors. It reflects the spoken content of a live webinar and has been lightly edited for clarity.

Aziz Al’Khafaji: Alright, good afternoon, everyone. It's a pleasure to be presenting at Roche’s workshop and share some of the efforts that we've embarked on with Mark and team over the past few months.

My name is Aziz Al’Khafaji. I'm the Director of Molecular R&D at Broad Clinical Labs and I lead the methods development lab, where we focus on creating and deploying cutting edge genomics tools.

When Roche reached out with the specs of the new SBX platform, we were excited to deploy it in ambitious ways that could mark a paradigm shift for key activities that we have going on at the Broad. We identified two main areas in the RNA space. The first is high-scale multiomic drug response profiling in multiplexed prism cell lines, and the second being ultra-high throughput single-cell RNA isoform sequencing with SBX. My main focus for this presentation will be on the single cell project, which Houlin and Guoping have led, with support from Brian. 

So just some very brief background. KRAS is a commonly mutated cancer driver and was previously considered an undruggable target, until researchers developed small molecule inhibitors which could inhibit that aberrant signaling. Despite very promising clinical outcomes, therapeutic evasion, drug resistance is still a big challenge. And if we can actually understand these routes, these mechanisms of evasion, we can actually drug those, have combination therapies, and realize the immense kind of potential within these drugs. So, rescue that potential.

To tackle this problem, we partnered with the PRISM team at Broad which has developed multiplex cancer cell lines consisting of 500 unique cocultured lines which can be distinguished through sequencing of their Express Prism barcode. This is an incredible resource which allows us to efficiently profile drug effects across diverse cancers in a single experiment.

So, we treated the PRISM pools with our RAS inhibitors BI-2865 and RMC-6236, along with our negative control DMSO, and a positive control that elicits a strong transcriptional response. So panobinostat, this is an HDAC inhibitor. We performed single cell at the three and six hour time points, and then as well a survival assay after five days to understand, you know, the survival outcomes.

With this experimental setup, we were able to actually link survival outcomes to early and intermediate kind of drug response. And the idea being, can we find these transient key intermediate states that allow for that ultimate survival, and those are potential targets that we can validate. And if proven true that we can go after them with secondary compounds.

And so, our primary readouts are the PRISM barcode. Again, this helps us assign and identify the specific cell line that the transcriptional profile is part of. We also, you know, this is a multiomic, single cell profiling. And so, we actually captured protein expression through the protein text discovery panel, which gives us reports on greater than 320 protein expression targets. This is in the context of the 10x Flex single cell platform which is a probe based, whole transcriptome read out on the Gem-X. And of course we're sequencing this on the SBX platform which gives us, for this experiment, 80 billion reads.

The data was phenomenal. I believe it was Mark who said that this picture looks like a picture from the Hubble telescope, and indeed this is a universe of data to explore. So I'm just going to go through some high-level metrics that we're all used to looking at with single cell. So here we had a median of about 1000 cells per cell line. For the per cell metrics we got about 28,000 UMIs per cell, 7,000 genes per cell. So the data was again very high quality.

There's no obvious batch effect detected. There's actually no green pixels on this image. This is actually the kind of co-clustering of our two different channels from single cells. So orange and blue kind of make green in a visual cortex.

Just as a high level sanity check, cell lines from the same tissues are more transcriptionally similar. So you can see on the bottom left the skin cancer cell lines, the top colon cancer and kidney cancer, etcetera. We also pseudo-bulked the different cancer cell lines from the single cell data, and compared that to orthogonally derived bulk RNASeq and they're highly correlated. So, we just have very strong confidence in the quality of this single cell RNAseq data.

Equally so, the protein expression data was very impressive. We were able to resolve many cell lines from protein expression data alone, which is pretty radical. So for the metrics there we see almost 13,000 protein UMIs per cell, and a median of 300 proteins detected per cell.

So, this slide underscores the utility of capturing both RNA and protein. So, some RNA protein expression is highly correlated, as you can see on the top right graph, and then on the bottom right graph some of it is poorly so.

So, each dot in that graph is a cell line. The X axis is RNA expression, the Y axis is protein expression from the panel. So you can see for a housekeeping like gene, it's highly concordant, right? But for things like transcription factors that are highly post translationally regulated, you know, i.e., protein degradation. If those switches are turned on and off, you're going to see a lot of difference there. 

So RNA is not always a good proxy for what's going on at the protein level, especially transcription factors. Obviously very important. So, this just highlights the power of capturing both of these modalities. The graph on the left just shows the range of correlation expression across cell lines. And so there's a lot of kind of, you know, that variance.

OK, let's dive into to drug response. As expected, we observed a range of transcriptional responses to our inhibitors. So annotating cells by treatment, we can use colocalization to get just a very high level, first glimpse at the various responses. Here we can see that the treatments are not colocalized with our negative control DMSO which is in purple, and this shows us that this is an example of a large transcriptional response to the inhibitors. Here DMSO and RMC treatments are colocalized so pointing to less of an effect on this cell line with the RMC drug. Vice versa for the BI drug, and here's a case where it looks like this line was not responding to either inhibitor.

OK, shifting focus now to the five-day survival data, where dark blue is sensitive and yellow denotes more resistant cell lines. I'll highlight four main classes that these cell lines fall into. Starting on the very left, this is resistant, cell lines that are resistant or sorry, sensitive, to both compounds. Just to the right, sensitive to RMC, furthest to the right, sensitive to BI only, and then in the middle resistance to both right.

In the interest of time, I'll just show that there are unique enrichments of mutation in these different classes, many of which are concordant with the literature. So, in the sensitive to both category, we can see that KRAS is actually enriched for, which is entirely in line with the mechanism of the drug. These drugs target mutated KRAS, and then we see different enrichments for various mutations for the uniquely sensitive cell lines, and then for the resistance of both, you can see that BRAF is one of the main enriched mutations, which makes total sense. This sits downstream of KRAS and does not need kind of this aberrant signaling from a mutated RAS to have its oncogenic effects.

So, this data was produced a little more than a month ago. So of course, this is just the beginning. From this first look, it's clear that there's a lot to discover from this data set. And we're excited to bring to bear, you know, these classical analysis, advanced AI approaches, and we will surface some of the complex biological signals.

And you know, we feel like this is a high yield approach. To systematically identify mechanisms of resistance, to be able to validate those, target them. And you know, I think really just mark a paradigm shift for drug development.

OK, great. Now I'll shift gears, and we'll briefly touch on our efforts in the RNA isoform sequencing space, with SBX. This work was led by Christophe, Brian and Houlin. So, we ran kind of like a traditional assay PBMCs and a 10x 5’ gene expression kit, and we sequenced the pre-fragmented folding cDNA libraries. We were able to resolve the canonical cell types and their biomarker genes, everything was kind of nicely clustered. But we wanted to actually dive in and, and see, you know, what these read lengths look like and their ability to resolve isoforms. So, on the far right is the read length distribution, the lion's share of which is between 500 and 800 base pairs. And you can see, you know, a fairly long tail heading to the right there.

To understand the impact of read length on our ability to resolve isoforms, we informatically binned those reads from 300 to 1200 in hundred base pair increments, and then we ran this kind of isoform completeness metric called, these categories called full splice match and incomplete splice match. We can see as we go from 300 base pair bin to the 500 base pair bin, there's a sharp increase in percentage of reads that are full splice match. So that's a good quality metric for isoforms. And so, this gave us a lot of confidence that these read lengths could span multiple splice junctions and even resolve full length isoforms.

And this is just a read level example of just that. We have the CD1C gene and we see those reads spanning the multiple exons. The bottom two are known isoforms and the top one you can see that second to the last exon missing. It's not an annotated isoform. So that's pretty exciting.

And so I'd like to acknowledge the immensely talented scientists in my group. The PRISM team, Broad clinical labs, and our industry partners. The single cell experiment in particular was an intensive effort and it took everyone coming together to make it a success. And it's been a sincere pleasure collaborating with Roche on this transformational technology. Mark, congratulations on this landmark achievement.