Scientist Stories: Evan Eichler, Long Read Sequencing of Complex Genomes

It's a pleasure to be here again at ESHG and especially at Barcelona. I'm going to be sharing with you today some of the advances that have been made, we think, with respect to long read sequencing, particularly from a focus on single molecule real-time sequencing. So first I'll start with a fairly general slide in terms of the kind of the tie that binds all of us in this room is really the pursuit of genetic variation and the linking to phenotype. That is essentially what human geneticists do and geneticists more broadly. And this is just to remind you that, in fact, there's a wide spectrum of genetic variation ranging from single base pairs all the way up to large chromosomal changes that we see under the light microscope. And our ability to characterize this variation has always been dependent upon technology. With cytogenetics 30, 40 years ago, with sequencing technologies now coming to the fore, allowing us to capture really single base pair mutations and other forms fairly routinely. Not all forms of genetic variation are created equal, and I'm going to be talking mainly about the application of long-read sequencing technologies to really fill in the spectrum of human genetic variation, really going from indels all the way up to large-scale structural changes defined here as insertions, deletions, and duplications within genomes. So the reason that detection of genetic variation is so difficult is our genome is complex. So this is just one of those complexities that is depicted here. What I'm showing you in this map is a schematic in blue of all the regions that are duplicated within a chromosome that are 95% identical and greater than 20 kilobases in size. And so you see this distribution of duplications. Some chromosomes are loaded with this duplication. And if you add the interchromosomal duplications on top of that, you get a pattern that looks like this. With accumulations of duplicated sequences near the ends and near the centromeres of chromosomes. So large swaths of our genome are made up of various classes of repetitive sequences, and in fact, in these areas of the genome, there are about 1,000 genes. There's more copy number variation in these regions of the genome, this 5%, than anywhere else in the eukaryote with respect to copy number variation. And yet our ability to assay this variation routinely or systematically is fairly limited. So this is the problem. Human genetics, we think we're good at detecting variation, and I would argue that we're pretty good at detecting large variants, so things greater than a few kilobases in size now, and in terms of detecting single nucleotide or point mutations within genomes for the vast majority of at least the euchromatin. But in fact, we're fairly poor for different classes of variation. So the fundamental limitation, I would argue right now in the field of human genetics, is our dependence upon short-read technologies and the fact that we map to discover. So we're taking short reads, and they're fantastic, they allow us to do it rapidly and cost-effectively, at least in a clinical setting, to discover variation, but not all forms of genetic variation are discovered equally as well. And the fundamental problems are are: mappability. So when you're talking about short reads like Illumina or even complete genomics, the sequences— reads are relatively short, and they don't allow you to map to complex repetitive regions, to general repeat regions, or even low-complexity regions with high accuracy. The technologies are limited. So in terms of their bias with respect to GC richness and AT richness. So many of you know when you sequence an Illumina genome, that areas that are particularly GC-rich or AT-rich are not well represented within those genome sequence libraries. And then there are biases against certain types of variants. So even the algorithms for calling insertions and deletions and structural variants— and we've been working on this area for many years— are highly imperfect and really dictated by low false discovery. So our ability to detect variants is often set by some standard where we want to keep the number of false discovery events relatively small, and as a result we throw out a large fraction of the variation. So this is just an example of a PCR-free genome that was sequenced with Illumina technology showing a particularly important locus called SHANK3. It's a risk factor for autism and developmental disability, and showing you the complete absence in a 40x sequence library over the promoter and first exon region. So you see essentially no reads over that area. And the reason for that, not a surprise, is that it's made up almost entirely of Gs and Cs. So this is a particularly kind of onerous example where we can't, despite no matter how much we sequence, really get high coverage over those areas. So long-read sequencing technologies is what I'm going to be talking about today, and I'm going to be specifically focused on PacBio. My disclaimer: I am not invested, I am not on the SAB, I have no money, I get nothing from this, from PacBio. So I'm presenting this data to you as a— really as a— from a perspective of a scientist interested in both basic and clinical research. So I plan to talk about both the strengths and the weaknesses of this technology. I want to talk about its application, specifically to understanding what I consider to be the missing genetic variation within human genomes, and how we can use this data coupled with Illumina data to more accurately genotype biologically and clinically relevant regions. I will end with some discussion about whole genome assembly, some of the first human genomes as well as some of the first non-human primate genomes that are being sequenced and assembled with this technology. So what is PacBio sequencing technology? Well, this is an instrument, this is actually a picture of the instrument that came to our lab about 4 years ago. It's a monster of a machine. It cost us about $1 million to get this instrument in and required about, I think it was 10 weeks of engineers to make sure that the configuration of the floor was correct. We had to find a good space that would actually be load-bearing for this monster instrument. The technology works on something called zero-mode waveguides. And so zero-mode waveguides are essentially channels about 50 nanometers in terms of their width, and there's 150,000 of them on a smart cell. And the way this technology works is that you in fact have your DNA polymerase This is an engineered DNA polymerase tethered to the bottom of the zero-mode waveguide, and you have, by diffusion, long DNA molecules, native DNA molecules, that diffuse into the actual zero-mode waveguide. And then you have phospholinked nucleotides that come in, and as they are being incorporated, they are excited by lasers and they emit fluorophores as they're being held at the bottom of that zero-mode waveguide. And they release signature phosphofluorescence which can be detected and in fact incorporated into a movie. And so the type of sequencing that actually goes on is envisioned here, is basically the polymerase at the bottom, phospholinked nucleotides come in, they're being held, as they're being held it results in a pulse which is being detected as a function of time from the actual lasers. It then becomes cleaved, it moves away from this area where it can actually be detected, and a new— becomes a DNA molecule, translocates, and a new phospholinked dinucleotide comes in, and a new pulse is initiated. And so the signature that you get is very different than what you're typically used to. But you imagine this ZMW is 150,000 of them, and now there's machines that are called Sequel that have on the order of a million. And you're doing this simultaneously. You're actually capturing DNA sequence in real time. So the strength of this technology, in part, is that there's no cloning, there's no amplification. I want you to remember that, because a lot of technologies that even involve bridge amplification from Illumina involve this idea that you have to create amplified material. You're starting with native DNA. The disadvantages are low throughput and a high error rate, around 15%, and most of the error is missed in corporations because the DNA polymerase moves so quickly sometimes you miss the actual incorporation itself. So the reads are long, as many of you know, and as the session title would suggest. The mean length that we get— this is a— I just pulled this off the machine before I came here to Barcelona. We're getting mean read lengths of about 12.4 kilobases, max about 88-kilobase sequence reads. This is a typical 6-hour movie in a distribution. What's most important for assembly and characterization of mammalian genomes is this: the proportion of reads that are over 20 kilobases. You need these super reads, we call them, or long reads to really initiate assembly and rediscover complex and anchor variation within complex areas of the genome. So it's not sufficient just to have a lot of reads. You need to actually have a significant fraction of very long reads. And I've defined that here as greater than 20 kilobases. Another really important feature of this technology is that it's in fact extremely uniform, and so what I'm showing you in the red dots is a human genome that's been sequenced with PacBio and its read depth for that genome. We targeted about 40x sequence coverage for this genome initially as a function of GC content. So what I've done is I've divided the human genome into bins, and these red dots indicate the coverage in various bins of 5% across the genome. What I'm showing you in contrast is 20 human genomes done PCR-free with Illumina. That's indicated in blue with these scatter plots as a function of GC content. And this is an important point. Although it's not perfect, you can see that the depression as a function of increasing GC content is only about 5%, whereas the case for Illumina sequencing, even with PCR-free, you're getting on the order of 50% reduction as a function of GC coverage. So what that means is that we can now access GC-rich regions of the genome uniformly, or at least as uniform as we could with, let's say, non-GC-rich portions. So coming back to our example of SHANK3, This is to show you PacBio sequence coverage for 40x compared to Illumina sequence coverage. I think you can see pretty clearly there's a big difference here in terms of the representation of this GC-rich. So now we can capture the promoter and the first exon and sequence and, in principle, assemble it. The last point I'll make about the technology is that if you get sufficient sequence coverage, you can, in fact, error correct. At pretty high accuracy. So this was developed by David Alexander, and essentially the idea is quite straightforward. It builds off of the idea that the error itself is fairly uniformly distributed. So I've looked at a lot of different sequencing technologies over the last 25 years, and I can tell you that in fact, in terms of the sequencing technologies, there's nothing that is totally random in terms of the error bias. But in terms of PacBio, it is probably the most random that I've seen. And so the idea behind getting high accuracy sequence is that if you generate high accuracy sequence, you can take the shorter reads and you can error correct the long reads, which are called super reads, and you can in fact develop an error probability model, which is what the algorithm Quiver in fact does, and compute an accuracy on the order of 1 error every 10,000 to 1 error every 1 million. 50,000 base pairs. So we've done this many times now, and we actually can see that the accuracy, once you get past 70x, gets up to that what we call a QV value of 30 to 40. So that means an error once every 10,000 or less in the sequence. But the key here is you gotta get coverage. And so this makes the technology expensive, because you can't be satisfied with 30, 40x. You gotta get to 50, 70x sequence coverage. Alright, so I want to show you data now in terms of application with respect to structural variation. So we did a project, and now it's about 2.5, 3 years ago, to sequence the first human genome with long-read sequencing technology. And it wasn't just any old human, it was a haploid human called the complete Heidelberg mole, and it is essentially not a human, but essentially a mass that grows as a result of a haploid kind of fertilization of an enucleated egg. So a sperm fertilizes an enucleated egg. It essentially creates a homozygous state for all chromosomes. You then actually generate the sequence data from that material. The reason for that is we wanted to actually build our first reference without the confusion of actually dealing with two genomes at the same time. So instead of having a maternal and a paternal complement, let's just start with one genome. So this was older technology. It was called P533. The newest technology is called P6C4. Its error was on the order of 15%, and its average read length was about 6.2 at the time we started this project. So we didn't try to set out to actually assemble the genome. What we wanted to do was actually discover structural variation, and we wanted to do it comprehensively. And the way we decided to do this was to take the reads initially and map them back against the human genome. What we would do is we would look for signatures where we would see multiple reads supporting a deletion, multiple reads supporting either an insertion or a deletion, or areas of the genome where all the reads would begin to align to the human reference and then they would stop. And no matter which way we did it, forward or reverse, we would actually see this hard stop. We also looked for signatures of the genome where we would align these long reads to the human reference and they would align, stop,, and then let's say 5 to 10 kilobases later they'd pick up again, and so we call that a double hard stop because there was something different over that stretch with respect to the human reference. And then we looked for reversals in order. So the idea was take long reads from a haploid source, map to the human reference, find the signatures of structural variation, grab those reads, do a local assembly, error correct with Quiver, so now that we have high accuracy, and then map out the breakpoints of structural variation. So this is what we did, and I'll just summarize the results. So using this technology, we were able to close or extend about half the gaps that remained in the human genome. So there were about 90 of the 100 and, I guess, 200 gaps that remained, at least at that time. Of those that we actually closed or extended, we could see that many of them were made of complex repeat structures, particularly short tandem repeats, various types of dinucleotides, triplet repeats, VNTR repeat sequences. We identified within that human haploid 26,000 structural variants, and 99% of these were resolved at the base pair level. So what does that mean? That means we were not inferring the presence of a structural variant based on read depth or read pair. We actually had the sequence structure of 99%. We knew the begin, the end, and the sequence structure of the structural variant. This included about 500 very large short tandem repeats and variable number of tandem repeats, 1,000 complex insertions, 36 inversions, and 2,600 sites which we call muted gaps, and these are gaps in the genome that you wouldn't recognize that are gaps because essentially the human genome has inadvertently corrected or closed this region of the genome but in fact should be open because there's additional sequence that in fact all humans have, it just wasn't present in the human reference sequence. And our validation rate, depending upon the class, ranged from 92 to 99%. So this, if you get any take-home message from this education session, this is it. When we go back and look at the structural variation, so we took the 26,000 and we said let's look at them as a function of size. Let's look at the insertions in black or gray with respect to the human reference. Let's look at the deletions in red or pink. And again, we're just blocking out here as a function of size. So everything that you see as black or gray is an insertion in this new human reference compared to the reference genome, and everything in red is a deletion. Everything that you see in gray or pink means that in the databases that have been generated for characterizing structural variation, this event has been seen before. So important point, you can see that 92% of the insertions and 60% of the deletions are novel. That is to say, despite our work and many other labs' work in trying to characterize the pattern of normal variation, we discovered essentially a large number of new structural variants in this new genome. For Alus and lines, which are 300 base pairs, and L1 events, which are 6 kilobases in size, you can see that that proportion is less, and that's because we know what we're looking for when we're detecting structural variations of this type. Of these 22,000 novel genetic variants, this corresponds to 11 megabases of sequence. So I want you to think about how much single nucleotide variation exists in this genome. This haploid genome has about 1.7 single nucleotide variants equivalent to 1.7 megabases of sequence. So when it comes to actually discovering variation at the base pair level, this is almost 9 times or 8 times more base pairs that are affected, but most of this variation isn't discovered by application of short-read sequencing technologies. And some of you will go, "Ah, it's just boring sequence." Well, if you care about areas of the genome that contain genes or coding sequences, or UTRs, you can see the breakdown. So 7,000 of the events map within 3,418 genes. These are the novel variants. And 170 map within the UTRs of genes, or untranslated regions. So a large fraction of genetic variation, despite the 1,000 Genomes and many other projects that have been led, are not discovering the pattern of variation, at least for this intermediate size. And you can see that most of the gain happens underneath 5 to 1 kilobase in size. So of the types of events that we saw, shown here are the insertions for short tandem repeats. We're plotting here the length in the new reference human genome compared to what we see in the— in this case, this is GRC37. You can see that the short tandem repeats are longer and more plentiful. So we see that the short tandem repeats that we document, This would include things like triplet repeats associated with genes and other things about 3-fold times more abundant and 2.8-fold more or longer in this reference, new reference compared to the human reference. And this just shows you the differential as a function of size. So when you get to events, short tandem repeats and VNTRs that are greater than 500 base pairs in size, there's a dramatic difference, a dramatic odds increase in terms of the types of variants that you're discovering. Using this new technology. Here's an example. So CGG repeats, I spent 4 years of my PhD career trying to clone and sequence this beast of a region of the genome, and we can pretty much rip through, sequence, and assemble it with PacBio. So incredibly GC-rich repeat, and if you know your CGG repeats in the human genome, you can see the pattern here. You can actually clearly see the AGG interruptions which is a nice healthy CGG repeat in a human genome, not having lost its repeats or its AGG interruptions. In terms of the muted gaps, so these are things that every human has but the human reference doesn't have. These are some of the examples of things, so anything shown here in the darker blue represents muted gaps. So these are things which every human will have but the reference does not, and the repeat content underneath it. And you can see that many of these muted gaps are just loaded with various classes of repetitive sequences, probably helping to explain why the human reference assembly, the Human Genome Project, didn't actually sequence and assemble these regions. And for inversions, you can detect inversions even at the single molecule level. So shown here are some examples of reversal orders within unique sequence in euchromatin. And this is just to show you validation of those inversions. Most of the inversions that we can detect with PacBio are less than 50 kilobases in size, and we know we're missing large inversions, particularly if they're flanked by extremely large segmental duplications, because the segmental duplications are larger than the read length that a PacBio can get through. So the critical part with a long read is that it needs to get from unique anchor to unique anchor in order to actually discover and characterize the event. And for those interested in centromeres and telomeres, we cannot assemble centromeres with this technology, but what we can do is extend into them and identify long reads. This is one which shows clearly the 171-base pair motif. It's a 20-kilobase read of alpha satellite DNA. We have extended into many of the centromeric regions, but we haven't gotten through them or assembled them. And the same goes for the ends of chromosomes. So shown here is the end of chromosome 14. There's this telomeric associated repeat sequence and then the T2A triple G sequence which is diagnostic of the end of a chromosome. So this technology or technologies like this that continue to build on read lengths I think are exciting because they'll get us to the point in this field where I think it will be possible to get a complete sequence and assembly of a chromosome from tip to tip. We're not there yet, and there are particularly problems, particularly around these super large repetitive regions of the genome. And I'll just make a shout out also for indels, even though this is not an area that we have focused on in our lab, is that with these long read sequencing technologies, we can now detect insertions and deletions, particularly where technologies like Illumina with GATK or Freebase drop off. So if many of you know, when you're looking at sequencing genomes, particularly clinical genomes, you don't get many calls that are believable, typically within above 10 base pairs in size. But with this technology, clearly we're getting lots of insertions and deletions, including we estimate about 500,000 in sequencing two different Huttetitiform moles. So these are two haploid humans that have been combined now to make a pseudodiploid to rediscover that spectrum of variation below 50 base pairs. And I've actually purposefully excluded 1 base pair insertion and deletion events because we know from the technology itself the most common error that we're getting is insertions and deletions of a single base pair because of missed incorporations. And in fact, we estimate just based on strict application of PacBio technology and quivering, you have an inflation for 1 base pair particularly one base pair events, of on the order of two-fold, suggesting artifact based on the technology itself. So once you've discovered this variation, one of the things that we've been actually very interested in doing is now that we have sequence resolution and we have the alternate allele, can we go back to the Illumina data that we've generated for tens of thousands of individuals, including patients, and can we actually genotype that variation even though it wasn't discovered with Illumina technology? And for many of you, you would know the answer already. The answer is yes. For a fraction of it, you can go back and you can actually pretty— we think pretty accurately genotype. So what I'm showing you here is an analysis of taking 16,000 structural variants that have been sequence resolved, going back to Illumina datasets. So each of these represent genomes that have now been sequenced at 30-fold or more with PCR-free Illumina technology and just asking a very simple question. Now that we know the sequence, let's say an insertion or deletion, can we actually now genotype it, presence or absence, within Illumina data? And you can see the CHM1 against— PacBio data against CHM1 Illumina. About 84% of the data can be detected and we think genotyped within Illumina data. And of the variants that we discovered, 66% of them are supported by 50% or more of Illumina samples. So we can detect presence or absence of that structural variant now that we have the alternate allele sequence. So this is just showing you some data. These were events that were not discovered by Illumina sequencing but can be genotyped. So they were discovered by PacBio. Sequence structure was resolved. So what I'm showing you here is a deletion of 3,913 base pairs that removes the DNase hypersensitivity site we think corresponds to an alternate promoter of this gene called SLC25A24, and we estimate is deleted in 9.8% of humans. And the way we can estimate that is by going back to the 1,000 Genomes data and just genotyping based on read depth. So this is no fancy genotyping, it's simply just going back and measuring read depth. And you can see that there are large— relatively large fractions of the human population that carry it. But interestingly enough, Europeans and African Americans are underrepresented for this particular deletion. It seems to be high in individuals of East Asian and South Asian descent. Here's another relatively large deletion that was missed at least by the 1000 Genomes Project. It removes the DNase I hypersensitivity site, and it clusters around the promoter of a gene called KDM4C. We've observed it deleted homozygously in about 0.4% of humans. And again, we use the alternate sequence structure to actually genotype precisely over the breakpoints of this region Using read depth from the 1000 Genomes Project, you can clearly see the individuals that are missing this promoter. So interestingly enough, in this case, it's seen in mixed Americans as well as Europeans, but essentially absent in Eurasians, Africans, and South Asians. So I'm pumped because I think that if we do this systematically, long-read sequencing of human genomes, we can fill out at least the spectrum of genetic variation between 2 base pairs to, let's say, 20 to 30 kilobases, the normal pattern. And we can actually understand, at least for common variants, we can understand the sequence structure of those and then go back to our patients to find potentially disease risk or disease-causing loci. So there's projects that have been initiated. I know there's many in Asia that are going forward on this. The U.S., as part of an extension of the 1000 Genomes Project, but also as part of a new reference genome project whose goal it is to build human references that are better than the current GRC38, which is a pretty tall order, have actually been launched. So two new human reference genomes have actually now been released. They're called— based on the haploid genome CHM1 and CHM13. And WashU and our lab has initiated new references for what we call continental references. So we're building a new— right now it's 1, but soon to be 5 new references from Africa. A new reference— 2 new references from— going to be from Amerindian populations. There's new references, about 4, that are coming from Asia. And some new references that we're trying to launch from different continental groups, including Melanesia and Indochina. As part of the 1000 Genomes Project, or at least the extension of it, we're also doing trio sequencing. So there are 3 trios, a Puerto Rican, an Asian, and a Yoruban, in which the data now has been generated and is publicly available. Deep sequencing on the child and about half X sequence coverage, so 50 on the child and about 20X on the parents for trios from the 1000 Genomes Project. And the mission is really twofold: develop more comprehensive structural variation that goes— at least understandings of genetic variation from really the 2-base pair level to the 20 to 30 kilobase level, and at the same time build new references where we're actually going to actually create entirely new references that are going to be representative for different continental groups. This is showing you some unpublished data.— and part of the reason I'm so excited. This represents here on the right work from our group and about 20 groups characterizing 2,500 genomes as part of the 1,000 Genomes Project for about 4 years. So a paper came out at the end of last year summarizing about 60,000 structural variants with high confidence. And shown here is the number of events that you see and the average size of events as as well as the median size of events. This is analysis of the first 3 trios, so these are 6 unrelated genomes, and the number of structural variants that we discovered and we think are high confident with an FDR probably on the order of less than 5%. So 100,000 structural variants, so doubling the number of structural variants that was discovered from the 1000 Genomes Project, and obviously pushing the threshold down to really, you know, tens of base pairs in terms of its size. So basically sequencing a few human genomes gives us much more structural variation than sequencing thousands of genomes with low coverage using short read. So the last part is essentially what I think is the holy grail, not just of long read sequencing but should be for this scientific community here, and that's to do routine de novo assembly. Because everything I talked to you about up to this point was still aligning to the reference. And if you remember what I said in the beginning, we need to move away from aligning to the reference genome. So there are a number of new assemblers that have been developed, one by Adam Phillippi called mHap. There's a new version called Canoe. There's another developed by Jason Chin called Falcon. And I'm going to tell you about Falcon because this is the one I understand at least the best and have the most data on. So what Falcon does, it takes essentially the shotgun sequence data that you've generated for a PacBio genome and focuses on the longest reads. They're called P-reads because they're basically going to be the reads upon which the assembly is built. And it uses the shorter reads to actually correct the longer reads. So you're taking the shortest reads, you're correcting the longer reads. And then you're focusing on that, those longer reads to build a string graph, which actually relates each of these based on the overlap, these long reads to each other. There's a third step which is part of the strength, I think, of Falcon, which allows you to actually bridge across repeats that are incredibly large by taking the longest read that bridges, it's called repeat unitig bridging, and then you actually generate a high-quality consensus using Quiver. For those of you interested, this assembly algorithm is available, and there's obviously datasets that are available now to use it. It is memory intensive, so you need a pretty strong compute farm to get it operating. It took us a month to get this running in our own lab. So if you apply the data from CHM1, which is the first human genome sequence with long reads, and you assemble it, you generate sequence contigs that are on the order of 25 to 30 megabases. Megabases in length. This N50 contig length means the point in which your 50% of your genome is in a contig of this size or greater. So 50% of the human genome is represented in contigs that are 30 megabases or greater by application of Falcon, and shown here are in black the alignments of these super contigs that have now been sequenced and assembled compared to the reference genome Everything greater than 3 megabases indicated in black. Anything indicated in green is less than 3 megabases in size. So you can assemble a human genome, at least a haploid genome, into 3,000 contigs representing on the order of 97% of the human euchromatin. Of course, that's a haploid. What happens if you actually do diploid genomes? The statistics aren't as good, but the data suggests that when you get to order of 80 fold sequence coverage, you can generate contigs routinely on the order of 10 megabases or greater. There's still a problem, and the problem comes back to those segmental duplications that I showed you in the very beginning. You remember those blue and red areas of the genome? If we ask the question, how good do these assemblers do within segmental duplication areas? I'm showing you both the Falcon assembler as well as mHap. And what I'm indicating here is the size of the segmental duplication, and then on this axis I'm showing you the percent identity of the segmental duplications. Anything in black means the segmental duplication was resolved. Anything in red means it wasn't resolved. And I think you can appreciate pretty clearly that once you get past about 50 kilobases or even let's say greater than 25 kilobases and you get to repeats that that are greater than 97% identity, you don't resolve those areas of the genome. And you can see this if you just take these Falcon and mHap assemblies, project them onto a UCSC browser, you look at where they're breaking apart, and you see that they're mapping right over the segmental duplication tracks. So segmental duplications cannot be resolved with— at least we can't yet. We're getting closer, but we're not there yet. And so we think in order to build higher quality references, 95, 96% of it can be done at least for the euchromatin with PacBio, and then what we're going to need to be able to do is go back in and actually target these areas of the genome using large insert clones, which is the way the Human Genome Project was done before. And so this is work that we're doing in the lab. We build BAC libraries for each of these human references. We're going back into the genome taking out the large insert clones, using a luminous sequencing technology to get the information content, and then picking tiling paths to resolve these areas of the genome. And this is just showing you an area which is an inversion polymorphism that you don't know about, or generally wouldn't know about, because it's completely embedded within segmental duplications. And yet, when we go ahead and actually generate an alternate library and sequence and assemble it, we actually can prove the presence of this 210-kilobase inversion totally contained within sigDUPs. And so this work, we've shown proof of principle of it, and now we're doing it systematically. We can go back to very complex areas of the genome that may be mediating rearrangements associated with congenital abnormalities, microdeletion syndromes, and so on. And we can resequence them by actually building a BAC library over the duplicated regions. And this is just an example where 32 BACs were selected over an area of the genome that wasn't resolved for 15 years. We targeted each of the BACs. We're now targeting groups of 4 or 5 within a SMRT cell, sequence and assemble, resolve, and in this case we corrected on the order of about a megabase of sequence, closing up most of the gaps and organizing the gene structure of those areas of the genome. So in the last few minutes I'll tell you about kind of a passion of mine, which is just to not only do additional human genomes and understand variation, but to do the great apes right. And I think I can say this without too much criticism because I've been involved pretty much in every great ape genome sequencing project, and I've been very unsatisfied with the quality of product that we've been generating, largely because we generate genomes that typically have on the order of 200,000 to 400,000 gaps when we apply short-read combined with Sanger-based sequencing technologies. And it's nobody's fault. It just simply is an issue of cost and time. So we've been going back and trying to now systematically resequence or to build new assemblies of the non-human primates. This is our first attempt. This is a gorilla genome assembly. Similar statistics as what you've seen before: 71-fold sequence coverage, N50, so kind of the average size of the contigs about 10 megabases, and importantly, we are closing most of the gaps of the genomes. So for those of you who don't know, the genome of gorilla up to about a year ago had about just shy of 500,000 gaps. So there were 433,000, and by doing sequencing assembly with PacBio, we closed 399,000 of those gaps. In terms of genes, we estimate that there are on the order of— on the order of about 4,000 genes that were incomplete. In the gorilla genome assembly. We now estimate that we've closed and got the complete gene structure for 88% of those gorilla gene models, and we've added 10,000 of the missing 12,000 exons into the genome. And this, I think, is important and probably has ramifications for why we're doing the work in human. If we go back to RNA-seq data or Illumina data that was generated for the gorilla. For RNA-Seq, we can now map back 8 to 9% more of the RNA-Seq data. And in fact, even in the eukaryotic regions of the genome, we're mapping on the order of 2 to 3% more of the Illumina data back to this new reference genome. And many complex regions that are not made up of high identity repeats can be sequenced and accurately resolved. So I'm just showing you here an example of the gorilla. It's not actually HLA, it's not— that stands for human leukocyte antigen. It's the MHC II region, and we can actually sequence and assemble the entire haplotype. In fact, we can actually tease apart the haplotypes of that particular sample, that particular gorilla. All right, so in summary, I've given you an introduction to PacBio sequencing technologies. Some of their strengths and weaknesses. Deep, smart whole genome shotgun sequencing provides much more comprehensive access to indels and intermediate size structural variation, particularly below 20 kilobases and all the way down to 2 base pairs in size. In fact, I would argue it's the first time we actually begin to see the full spectrum of that type of variation. 80% of the structural variation that we now know exists within genomes cannot be characterized or discovered by Illumina sequencing-based approaches, at least if you keep a modest FDR. But the good news is, is that once you've sequence resolved it, you can genotype it. We believe at least 65% of it can be genotyped accurately. New human references, new non-human primate genomes can now be sequenced and assembled with the long-read technologies. And we're routinely getting contig lengths greater than 10 megabases. The segmental duplication regions of the genome, the heterochromatic regions, still are problematic, but I believe with further improvements in both the technology, we should be able to capture and resolve those regions of the genome. So perhaps the final point I'll make is that I think the technology here, and it doesn't have to be PacBio, but the idea of long-read uniform sequencing technology is the future of human genetics. As a field, we need to move away from aligning sequences to discover genetic variation. We need to think about not just assembling a new genome to discover the full spectrum of genetic variation, because that's obviously what patients and what basic researchers are required. But we need to move to the day where we think about assembling not a 3-gigabase genome, but a 6-gigabase genome, where we fully resolve both the maternal and paternal complement And this is actually happening, and we can see this happening with this technology combined with 10X and other technologies coming together. So this is the future, and I think as a field— and I've said this a couple times, but I'll just say it again— if we're still aligning reads to genomes to discover variation, I think we've failed. I think we need to be thinking about de novo assemblies of genomes for the purpose of basic research as well as disease. And these are the folks whose work I largely presented. I'm happy to take questions.