Scientist Stories: Euan Ashley, The Potential of Ultrarapid Nanopore Genome Sequencing for Medicine

I've been showing a picture of this Ferrari for the last 10 years as I give talks to various audiences around the world about genomic medicine. I used to use it in place of the Moore's Law graph that we're all probably pretty familiar with to persuade people that the cost of human genome sequencing was dropping dramatically in price.. And I was driving home one day through Atherton in billionaire territory in Silicon Valley. And I scanned to my left where the Maserati Ferrari garage is. And I saw this car. A little bit expensive even for my Stanford salary. But I calculated the math in my head at that moment, which is to say that if this Ferrari had dropped in price as much as human genome sequencing had dropped in price, then instead of $400,000, you'd be able to get this Ferrari for 1 cent. And over the course of the last 10 years, our community in industry and academia has moved towards rapidly bringing down the cost of human genome sequencing for discovery science and for clinical medicine. We've also, as well as reducing the cost, focused carefully on increasing the accuracy. And this is something that we, in particular on the clinical side, have really focused on for all the technologies, to say if we're going to start making patient decisions with these technologies, then we're really going to have to make sure we have great confidence in the variants that we're calling. But sometimes it's not enough to simply have a low-cost genome and a highly accurate genome. There are some situations in medicine where speed is important. I'd like to introduce you to Matthew Kunzman. This Matthew is a patient who came to Stanford Hospital just a few, a couple years ago, basically through the winter of that particular time. He'd been noticing that he had a cough and he wasn't feeling particularly well. He had a couple of weeks history, pretty short history in total, and so went to his local hospital. The assumption of both his parents and him was that he had COVID. Of course, of course you would at that point in the first winter of COVID that we had. So he was admitted to hospital in Oregon where where he had an ultrasound of his heart and actually was diagnosed with heart failure. And so it turned out that in fact this wasn't a respiratory disease, but rather one related to his heart. And this is actually his heart shown here. This is a cross-sectional picture of his heart. It's moving, as you can see, but it's not moving very much. A normal heart will push out 60% of the blood per beat. You can see this heart is barely managing to push out 10% of the blood per beat. And in fact, all of that blood isn't even going forward. So this is now a picture of the heart from the bottom. You can see the color of the blood moving, and the blue jet that is going upwards towards the top of the screen indicates that a significant portion of the blood that his heart is beating is actually going backwards. So he was in a condition called cardiogenic shock, and he was airlifted to Stanford Children's Hospital. Cardiogenic shock is where your heart is essentially not providing enough blood and oxygen for the needs of your tissues. He was crashed onto something called ECMO, extracorporeal membrane oxygenation. This is a heart-lung machine that basically takes over the pumping of his heart to keep him alive while we were able to organize an operating room for him to have a surgical implantation of an artificial heart. In fact, if you look up at the picture of him just now, you'll see the actual artificial heart that was placed in his body. And that photograph of him there You'll notice he has a sort of side pack, and that's the battery for his artificial heart. So this is a picture taken of him as he was recovering from that initial admission. But what can cause your heart to become weak like that in a very short period of time? Not many things. One of them is myocarditis, an inflammation of the heart muscle. And if that was the case, then you'd expect to find infiltrating lymphocytes in the heart tissue. But of course, It's also possible that there was a genetic cause. And this was critical because in his case, if it was myocarditis, there's a chance that he might recover completely. So he could have this artificial heart removed, his own heart could be sewed back up, and he could recover and essentially live a normal life. That happens not uncommonly with myocarditis. If, on the other hand, he had a genetic cause for his heart condition, this was going to be irreversible, and he should be listed as soon as possible for a heart transplant. So what would we do? Well, the normal way of making this transplant decision would be to do a biopsy. Take this medieval-looking claw that you see here at the top of the screen, pass it through a vein into his heart, and basically grab pieces of his heart muscle. And we did that because that's the standard of care. In fact, that's a piece of his heart shown at the bottom there with the trichrome stain showing the fibrosis in blue. It did not show any lymphocytes, but you can only take a piece of the heart from a couple of little places. So imagine if there was a way to understand whether there was a genetic basis for his heart condition in a way that we could pass that back to the critical care team. But genomes take too long, don't they? Well, that is in general what we have thought. In 2012, Stephen Kingsmore, at the time at Kansas, now at Rady Children's, started a legacy of 10 years of thinking about whether genome sequencing could be done faster, not just for not for clinical care, but for critical care. And working on the Illumina platform in 2012, you had this landmark paper showing that if you put one sequencing machine, at the time a HiSeq, entirely aside for one genome, you could sequence it around 50 hours. We worked with the same Illumina team at the time to knock 8 hours off the computational pipeline, publishing in these two papers shown here, two neonates who both presented with a cardiac condition of a heart rhythm abnormality that made their heart stop at the early part of their life called Long QT syndrome. And over the course of the next 10 years, Stephen managed to push that down. In fact, at 26 hours, he was awarded a Guinness World Record for the fastest genomic diagnosis on this platform, and then beat his own record just a few years later at 19.5 hours. So clearly it's possible to get genetic information faster. But in reality, in 2021, even in the US, the advertised clinical test turnaround times are not anywhere close to that. So that's the fastest you can get, but it's not the routine by any means. In fact, 6 to 8 weeks is about how long it's going to take to get a clinical genome test report in 2021. If you order a rapid one from Baylor, you'll get it back in 7 days. If you order a rapid one from, from Rady, where Stephen's group carries out this clinical genome sequencing, then it'll still take 5 days to get a provisional report. Patients like Matthew don't have 5 days. They need answers much faster than that. So I was having this thought one night: can we go faster? Sitting at home thinking, surely we have the technology now to go faster than that in a way that would be relevant. Can we do it also by maintaining or even improving the reach and accuracy of our platform? Well, maybe we could. And so that's the idea that I pitched to John Gorzynski, who's here in the audience today, and you can hear a talk from him later. Would it be possible to do this faster? And he said, "Yeah, I think it probably would be possible to do this faster." I don't think any of us really believed we could make it as faster as in the end we were able to do, and as I'll explain to you over the course of the next few days. The other students led this incredible team, and you'll hear more about all the different components over the next few minutes. But my first call was to our friends, of course, at Oxford Nanopore. Because what technology might be capable of outputting the sort of data that you'd need to assemble a genome in a shorter time? Well, it seemed to me that there was really only one. And in fact, this machine, the Prometheus machine, with 500,000 reliably informed pores in one machine, is capable of putting out a 1x genome in 1 to 2 minutes. We're talking about genomes in minutes. That's the sort of speed that you need if you're going to try and apply this to critical care. Medicine. So we talked to the team at Oxford Nanopore locally in San Francisco, and of course the folks here, here in the UK and Oxford, Gordon and everyone else, were enthusiastic to collaborate. A little wary because it sounded a little bit pie in the sky at the time, but we were confident we could get the people together and then give this a shot. Could we actually go faster? And the people we had in mind were those who'd been building really the best tools for interpreting this data and with the highest accuracy. Thinking about the Genome in a Bottle tooth competition from the National Institute of Standards and Technology, we were really eager to work with the combination team from the University of California, Santa Cruz, Benedict Patton's group, and Andrew Carroll's group at Google Health. We needed to be able to call not just small variants, but, of course, structural variants, so Fritz Sedlaczek at Baylor joined the team as well, and through the ParaBricks implementation of those tools, we also pulled NVIDIA into the the team. And everyone came together. We started meeting on a weekly basis to see, would this actually just be possible? And let me give you a sense of the ultra-rapid workflow that we put together. We really aimed to look at every single part of the workflow and say, can we make it faster? And John, who I showed you in the earlier slide, is actually an ultra-distance runner. And so we used to joke that he was going to start by running the blood sample from the bedside of the patient to the lab. I think he actually did that. At one point. The fact he was a competitive runner is perhaps one of the reasons why we— our competitive spirits were raised with the idea of going faster. So addressing from the first point of the sample, just thinking about how to maximize speed while we continue to maximize quality, both of course of size and of DNA quality so that the library preps would be of sufficient quality. We weren't going to do 48 library preps, but it turned out that he could spread 8 library preps over the course of the entire Promethean, and that that would work exceptionally well. One of the biggest challenges— I'll tell you a little bit more in a moment— was the compute with this much data coming off one machine. So we'll talk about that in a little bit more detail, but the basic computational pipeline is shown here. For base calling, we used the standard Guppy approach, of course, which is optimized for GPUs. The minimap aligner based on the Burrows-Wheeler Heuristic from Heng Li, of course, pretty standard for long reads, but obviously also optimized really for CPUs. The team from the Pepper-Margin-DeepVariant group from UCSC and Google, of course, that is a combination of different algorithms that include neural networks like an RNN. For Pepper, DeepVariant's a convolutional neural network, along with Margin, which is a hidden Markov model for doing haplotyping on this whole genome data. So you can see that each of these computational processes really requires a different optimization structure. And then we think about the predominant split-read approach from SNFLs. And it really requires something bespoke, and so this is what the entire team came together to try to design. The variant prioritization was part of our clinical genome approach at Stanford, so we adopted that and adjusted it a little for the specific criterion of split-reads. Speed. So let me take you through how long each part took for Matthew's genome, because we recruited Matthew through his parents into the study. So the first part, as I mentioned, the library prep, the DNA extraction, was the modified Qiagen PureGene approach fragmentation. Pretty standard. That's 19 minutes. The library prep, which started with including barcoding, was 171 minutes. So let's put that up against the record, shall we? So here it is. 4.2 hours so far, the world record, I remind you. At the time we were doing this was 19.5 hours. So then we get to the sequencing part. So 200 gigabases of data, 1.26 per minute, N50 was 22, a depth of 63x. So this is the genome that we were putting together. Let's put that up there. Well, now we're at 6.9 hours heading for the world record of 19.5, pretty much on track at that point. But let's take a short diversion just to talk about how we deal with the computation. And this is the work of Sneha Goenka, who's also here and presenting later on if you want to hear more or get more details. Because with the compute power that came at the time with the Prometheion, if you wanted to put an entire genome, which could be the data from which could be developed over 1 to 2 hours, through that compute power for alignment and calling, that was going to take 18 more hours. And that really, that kind of time we didn't have. So we, Sneha, set out to try to fix that. And we decided to use a cloud method essentially where we could essentially optimize the pipeline for the different processes that were required and parallelize individual processes so they could happen essentially in real time. So transferring the data up to the cloud was the first part, reducing the file size and having these every 3 minutes basically having data pulled up and then moved across the different processes so that they could be, the instances could be optimized in the different ways for the different algorithms that we had. So as you can see here, for example, the GPUs, for instance, were 4, CPU instances 4 to 8, and the total instances on the go at any one time would be 16. So those instances would essentially be pooling data in the time to allow base calling at the same time as variant calling was happening. If you put all that together, I'm sorry to go through it so quickly, You basically move from a point where we have an overhead of 16 hours after sequencing to a point where the overhead is less than 30 minutes. So this is essentially real-time base calling and real-time alignment. We pass it on to the variant calling then, and here's back to Matthew now where you can see where we are. Here's the base calling, here's the alignment, this is the real data, and here's the actual variants. So 4.5 million variants as expected, 20 structural variants out of sniffles, that's 3.6 hours, so let's add that to the top. We get to 10.5 hours after the blood sample. The variant pipeline for prioritization, this is essentially a version of filtration. We're taking standard approaches to population databases, to existing disease databases. We use the Human Phenotype Ontology terms to be able to filter these down in order to take us from the 4.5 million variants you see at the top down to a number that is manageable for manual curation, because the last mile is still manual. What we found, though, was that right at the top of that list of prioritization was this heterozygous duplication in troponin T. And troponin T is a gene that is specific to the heart and is known to cause cardiomyopathy. And so we were able to gather a team almost immediately and make the diagnosis and pass it back to the team at the bedside. The time on that was 11.25 hours, 8 hours faster than the fastest time that had ever been carried out. We confirmed this in our clinical lab through a very quick Sanger and then moved forward to the point where Matthew was urgently listed for heart transplantation because we had shown this could not be myocarditis. He carried a significant variant in the troponin T gene that explained his cardiomyopathy. It was irreversible. Here's a picture of him smiling in the first 24 hours after he had his heart transplant. And here is a picture of him just a few months later standing next to the machine that sequenced his genome at that critical moment in his path. The opportunity to affect the lives of individuals through this kind of technology is a real privilege. And we felt really excited to be able to bring this team together and apply this technology to multiple patients. Here's the first 7 patients, the first set of flow cells that we used, and you can see for each of the different steps where exactly the time was spent. You can see it's highly variable on the curation side. If we see a variant pretty quickly, it can be pretty fast. Otherwise, it can take a little bit longer. One of the ways that we were looking even within the study to speed things up was to consider whether barcoding was really necessary. Because we can— we do a nuclease wash as we reuse the flow cells. The question is, do we really need to do the barcode? In particular because we noticed that the poor occupancy, shown in the middle here, and the actual amount of DNA that needs to be loaded was actually significantly greater when we used the barcoding protocol. Now, so the gray bars, the ones with the gray at the back, are a new set of flow cells. So if you want to compare the difference with barcoding, you want to look at the blue and the red just immediately before that gray zone. What you can see is that the pore occupancy is significantly greater without the barcoding, and the amount of DNA loading is also higher. The question is, does that change accuracy in any way? And so this, again, work from Kishwar Safin, shown here on the HG002 genome from the Personal Genomes Program and from the Genome in a Bottle group. Basically demonstrating that the F1 scores for recall and precision were identical even without the barcode. So the small amount of carryover that there was was essentially not changing our accuracy scores in any way. And in fact, he also did analysis modeling that to show that even if the carryover was 10 times what we were actually getting, it still wouldn't affect these numbers. So with that, we had confidence to move forward for the next group of flow cells and basically remove the barcoding step.. And doing that, what we were able to do was essentially run each one of those genomes in under 12 hours. So in this particular pilot, every single genome diagnosis that was made— and the ones that were diagnoses you can see shown here in green on the left— every single one of these genomes was carried out faster than the previous world record for the fastest genome using this technology. What were the patients presenting with? Well, it's a little small. Shown here, but you can see it's a combination of neurodegenerative— neurodevelopmental disorders, seizure disorders, not uncommon, cardiac critical care, cardiac arrest, rhythm disorders. When I talked to the head of our neonatal intensive care unit, he says he would like to sequence two-thirds of the patients. And much of the work that Stephen Kingsmore has done over the course of the last 10 years has shown that about that number of patients would benefit if we deployed this technology at scale. And the benefit directly related to the fact that they can spend less time in the ICU. So it costs about $10,000 a day to have a patient in a critical care unit. So for a test that might be just a few thousand dollars, if you can save the healthcare system money— this is a test that can save the healthcare system money. Most new technologies cost money, and you have to decide, is it worth it? Here's a test that can actually save the healthcare system money. And what about our new approach then compared to the current standard of care? 6 to 8 weeks, there's the Baylor and radio ones that I showed you at the beginning. Here is the new ultra-rapid one on the same curve. Really potentially game-changing technology for these patients who are at their most critical moments. Compared to history, if we plot the new genomes here, just shown here, this is from the Human Genome Project and our own little version of Moore's Law here. And you can see the Guinness World Record there and the new genomes. Notably, these, of course, are long-read genomes, so there's much more information potentially available as well compared to the other approach. We summarized the work in these three papers. If you're interested to dig in a bit deeper, then please feel free to take a look at those. And as was mentioned in our introduction, having beaten the Guinness World Record, we were happy to be able to to be awarded a new Guinness World Record for the fastest DNA sequencing technique. It's a huge team effort. This is the first time the team were actually able to come together. We'd otherwise mostly met on Zoom. And this was the Guinness World Record ceremony next to our big DNA balloon sculpture. This is the sort of work that takes time. I came into my office one day and found this bag. If you look a bit more closely at the sticker, You'll see this says John's sleeping bag. And I find that my office has John's sleeping bag in it because literally we were at times sequencing through the night so that the team could get the data first thing in the morning. We're really going to have to rethink how we do genomics if we start to move from thinking about weeks and months to moving to start thinking about minutes and hours. And this is what it looks like when you make a diagnosis late at night on a Zoom call. This is us. I think, with one of the world record diagnoses. And this is often the way things happen, sort of late at night when we thought we had the answer. I couldn't leave you, of course, without bringing you back to the Ferrari. Well, what if we made this Ferrari as much faster as we've made human genome sequencing faster? Well, its top speed is 199 miles per hour. Well, if we made it as much faster as we just made human genome sequencing faster, it would go at 14 million meters per second. I was hoping that was warp speed, but it turns out that it's still 1/20th the speed of light. So we still have some work to go. Massive team effort. I cannot tell you, and I cannot say enough about just the amount of dedication and work that went into this from all these groups across multiple different companies. Of course, massive thanks to the Oxford Nanopore team in particular, many of whom are here with us today. It could not have happened without this massive collaborative effort. Thank you to them. And if you like these kind of genome stories, as was mentioned in the intro, feel free to take a look. There's a few more of them, including our early attempts at the fastest genome in this book. Thanks so much for being here. Thanks for joining me.